New Features and Improvements:
   - Added a --split-train-test option to imglab.
   - Added a max runtime option to the oca solver and structural object
     detection trainer.
   - Added an option to do translational jittering of the bounding boxes in the
     shape_predictor_trainer.  Using this option allows you to train
     shape_predictors that are somewhat more robust against variation in
     detection box placement. 
   - Added an optional thread_pool argument to find_max_global() and
     find_min_global() to allow for parallel execution of objective function calls.
   - Added auto_train_rbf_classifier(), a tool that trains a RBF-SVM without requiring the user
     to supply any hyperparameters.
   - Added polygon_area()
   - The image view objects now implement the generic image interface.
   - Updates to the Python API:
       - Added Python bindings for: extract_image_chip(),
         extract_image_chips(), center(), get_histogram(), sub_image(),
         polygon_area(), auto_train_rbf_classifier(), reduced(),
         translate_rect(), spatially_filter_image(),
         spatially_filter_image_separable(), num_separable_filters(),
         threshold_filter_singular_values(), max_point(),
         max_point_interpolated(), zero_border_pixels().  Also added an option
         for training nuclear norm regularized HOG detectors.
       - Added a bunch of new overloads and operators for existing Python types.
       - Made decision functions more viewable from python.  You can now inspect
         the contents of the functions rather than them being completely opaque.
       - Made it so you can call the std::vector<detectors> version of
         the object_detector constructor from python. So now you can pack
         multiple detectors into one object via the python API.
       - Allow batched face recognition for greater performance.

Non-Backwards Compatible Changes:

Bug fixes:
   - Fixed the point and dpoint operator + doing subtraction in the Python API.
   - Fixed label_connected_blobs_watershed() not giving contiguous labels as outputs. 
   - Fixed AVX detection not working correctly on some systems.
   - Fixed dlib.threshold_image() ignoring the optional thresh argument in the Python API.
   - Fixed hough_transform::find_pixels_voting_for_lines(). It would sometimes
     include a pixel multiple times in the output lists.
   - Made the routines that rotate image datasets use the rectangle_transform
     instead of the bad old way that only really worked for square boxes.  This
     improves placement of such rotated rectangular boxes.

New Features and Improvements:
   - Added a lot of new Python bindings.  You can now use these things from Python:
      - gaussian_blur(), label_connected_blobs(), randomly_color_image(), jet(),
        skeleton(), find_line_endpoints(), get_rect(), shrink_rect(), grow_rect(),
        image_gradients, label_connected_blobs_watershed(), convert_image(),
        convert_image_scaled(), dpoint, centered_rect(), centered_rects(), length(),
        as_grayscale(), pyramid_down, find_bright_keypoints(), find_bright_lines(),
        find_dark_lines(), find_dark_keypoints(), suppress_non_maximum_edges(),
        find_peaks(), hysteresis_threshold(), sobel_edge_detector(), equalize_histogram(),
        resize_image(), hough_transform, remove_incoherent_edge_pixels(),
        normalize_image_gradients(), line, signed_distance_to_line(), distance_to_line(),
        reverse(), intersect(), count_points_on_side_of_line(),
        count_points_between_lines(), dot(), normalize(), point_transform_projective,
        find_projective_transform(), inv(), transform_image(), angle_between_lines(),
        extract_image_4points(), load_grayscale_image(), min_barrier_distance(). 
      - Added a .add_overlay_circle() to dlib.image_window.  Also made .add_overlay()
        take lines.
      - Added the *_corner() routines to rectangle and drectangle and made these 
        objects constructable from instances of each other.
   - Made the Python extension module automatically enable AVX instructions if the host
     machine supports them.  So you never need to say --yes USE_AVX_INSTRUCTIONS anymore
     when installing dlib.

   - New C++ routines:
      - Added an image_window::add_overlay() overload for line object.
      - Added angle_between_lines()
      - Added extract_image_4points()
      - Added is_convex_quadrilateral(), find_convex_quadrilateral(), and no_convex_quadrilateral.
      - Added python_list_to_array()
      - Added min_barrier_distance() 

Non-Backwards Compatible Changes:

Bug fixes:
   - Fixed numpy_image and pybind11 crashing python sometimes when certain types of
     conversions are attempted.
   - Fixed some python functions not taking as wide a range of image types as they did in
     previous dlib versions.

New Features and Improvements:
   - Added Python interface to threshold_image() and partition_pixels().

Non-Backwards Compatible Changes:
   - In the Python API, renamed dlib.save_rgb_image() to dlib.save_image().

Bug fixes:
   - Dlib 19.11 had a bug that caused the Python interface to reject grayscale images.
     This has been fixed.

New Features and Improvements:
   - Deep Learning
      - Added resize_to layer.
      - Made loss_multiclass_log_per_pixel use CUDA. This greatly accelerates 
        training with this loss layer.
   - Image Processing
      - Added normalize_image_gradients() and remove_incoherent_edge_pixels().
      - Added neighbors_24 (for use with label_connected_blobs())
      - Added partition_pixels() and made threshold_image() use it to find the
        default threshold if none is given. Also depreciated
        auto_threshold_image() since using partition_pixels() to pick the
        threshold is superior.
      - Added overload of hysteresis_threshold() that uses partition_pixels()
        to select thresholds.
      - Added encode_8_pixel_neighbors() and find_line_endpoints().
      - Added image_gradients, a tool for computing multi-scale first and
        second order image gradients. 
      - Added find_bright_lines(), find_dark_lines(), find_bright_keypoints(),
        and find_dark_keypoints(). 
      - Added label_connected_blobs_watershed().
      - Added find_peaks().
      - Added find_pixels_voting_for_lines(), find_strong_hough_points(),
        and perform_generic_hough_transform() to the hough_transform.
   - Improved the stopping condition of the solve_qp_box_constrained() and
     solve_qp_box_constrained_blockdiag() quadratic program solvers.
   - Made find_min_single_variable() more numerically stable.
   - Make visual studio build with all cores when building dlib.
   - Added steal_memory() to the matrix.
   - Improved the workflow for adding part annotations in imglab.  You can now
     use shift+click to rapidly add parts to an object.  Each part will
     automatically receive an integer label.
   - Added .begin() and .end() to array2d.
   - Added the line class and some related utility functions.
   - Added centered_rects().
   - Added numpy_image, which is a simple type safe interface to numpy arrays.
     It is also a fully functional dlib image class, making interfacing dlib's
     image processing to Python much cleaner.  

Non-Backwards Compatible Changes:
   - All CUDA code in dlib was moved to dlib/cuda.

Bug fixes:
   - Replaced all the old Python C APIs interfacing to numpy with the new
     numpy_image.  This fixed some reference counting errors as a result.
   - Fixed load_bmp() not loading certain types of BMP images. 
   - Fixed compile time bugs when trying to use CBLAS in some unusual environments.
   - Fixed bug in global_function_search's constructor that takes initial function
     evaluations. It wasn't assigning these values into the entire state of the
     solver.
   - Fixed compile time error in the matlab bindings.  Also, fixed cell arrays
     of complex types not binding correctly.
   - Renamed BOOST_JOIN to DLIB_BOOST_JOIN to prevent name clashes when working
     with boost.
   - Fixed bug in hysteresis_threshold() that could cause incorrect output.
   - The last release of dlib added more ODR violation checks.  However, these
     could erroneously trigger when using Visual Studio and CUDA in certain
     workflows.  This has been fixed.
   - Fixed matrix objects not working correctly when sized greater than 4GB in
     Visual Studio.

New Features and Improvements:
   - Deep Learning:
      - Added scale_ layer, allowing implementation of squeeze-and-excitation networks.
      - Added loss_multimulticlass_log: used for learning a collection of multi-class classifiers.
   - Added a random forest regression tool. See random_forest_regression_trainer.
   - Added make_bounding_box_regression_training_data()
   - Added isotonic_regression
   - Added momentum_filter, rect_filter, find_optimal_momentum_filter(), and
     find_optimal_rect_filter().
   - Added binomial_random_vars_are_different() and event_correlation().
   - Added xcorr_fft(), a routine for efficiently performing large cross-correlations using the FFT.
   - Added the ramdump type decorator for invoking faster serialization routines.
   - Added check_serialized_version()
   - Added max_scoring_element() and min_scoring_element()
   - Made orthogonalize() faster.
   - Updates to the Python API:
      - Added interface to the global_function_search object.  This is a more general
        interface to the solver used by find_max_global().
      - Added support for variadic Python functions in find_max_global().
      - Added rect_filter and find_optimal_rect_filter().
      - Added make_bounding_box_regression_training_data()
      - Added the image_dataset_metadata routines for parsing XML datasets.
      - Added rvm_trainer
      - Added probability_that_sequence_is_increasing() 
      - Added dlib.__time_compiled__ field
      - Added num_threads to shape_predictor_training_options.
      - Added CUDA controlling routines such as set_device() and 
        set_dnn_prefer_smallest_algorithms().

Non-Backwards Compatible Changes:
   - Changed CMake so that there is only the dlib target and it isn't forced to
     be static.  Instead, the build type will toggle based on the state of CMake's
     BUILD_SHARED_LIBS variable.  So there is no longer a dlib_shared target.
   - Changed the integer types used to represent sizes from 32bits to 64bits in numerous
     places, such as in the tensor object.  This should be a backwards compatible change
     for nearly all client code.

Bug fixes:
   - Fixed memory leak in java swig array binding tool.
   - Fixed windows include order problem in all/source.cpp file.
   - Fixed cont_ layers not printing the correct num_filters parameter when they were 
     printed to std::cout or to XML.
   - Fixed some code not handling OBJECT_PART_NOT_PRESENT values correctly.
   - Fixed fft_inplace() not compiling for compile time sized matrices.
   - The shape_predictor_trainer could have very bad runtime for some really
     bad parameter settings.  This has been fixed and also warning messages about
     really bad training data or parameters have been added.
   - Fixed the decayed running stats objects so they use unbiased estimators.

New Features and Improvements:
   - Switched the Python API from Boost.Python to pybind11.  This means Python
     users don't need to install Boost anymore, making building dlib's Python API
     much easier.
   - Made the sparse version of svd_fast() use multiple CPU cores.
   - Changed the behavior of imglab's --flip option.  It will now attempt to
     adjust any object part labels so that the flipped dataset has the same
     average part layout as the source dataset. There is also a new --flip-basic 
     option that behaves like the old --flip. However, most people flipping a
     dataset with part annotations will want to use --flip.  For more details
     see: http://blog.dlib.net/2018/01/correctly-mirroring-datasets.html

Non-Backwards Compatible Changes:
   - Removed std::auto_ptr from dlib's old (and depreciated) smart pointers. 

Bug fixes:
   - Fixed global_optimization.py not working in Python 3.

New Features and Improvements:
   - Added a global optimizer, find_max_global(), which is suitable for
     optimizing expensive functions with many local optima.  For example, you
     can use it for hyperparameter optimization.  See model_selection_ex.cpp
     for an example.
   - Updates to the deep learning tooling:
      - Added semantic segmentation examples: dnn_semantic_segmentation_ex.cpp
        and dnn_semantic_segmentation_train_ex.cpp
      - New layers: loss_ranking, loss_epsilon_insensitive, softmax_all, and loss_dot.
      - Made log loss layers more numerically stable.
      - Upgraded the con layer so you can set the number of rows or columns to
        0 in the layer specification. Doing this means "make the filter cover
        the whole input image dimension".  This provides an easy way to make a
        filter sized so it will have one output along that dimension,
        effectively making it like a fully connected layer operating on a row
        or column.
      - Added support for non-scale-invariant MMOD.
      - Added an optional parameter to dnn_trainer::get_net() that allows you
        to call the function without forcing a state flush to disk.
      - Sometimes the loss_mmod layer could experience excessively long runtime
        during early training iterations.  This has been optimized and is now
        much faster.
      - Optimized the tensor's management of GPU memory.  It now uses less memory
        in some cases.  It will also not perform a reallocation if resized to a
        smaller size.  Instead, tensors now behave like std::vector in that
        they just change their nominal size but keep the same memory, only
        reallocating if they are resized to something larger than their
        underlying memory block. This change makes some uses of dlib faster, in
        particular, running networks on a large set of images of differing
        sizes will now run faster since there won't be any GPU reallocations,
        which are notoriously slow.
      - Upgraded the input layer so you can give
        input<std::array<matrix<T>,K>> types as input. Doing
        this will create input tensors with K channels.
   - Added disjoint_subsets_sized
   - Added Python APIs: get_face_chips(), count_steps_without_decrease(),
     count_steps_without_decrease_robust(), and jitter_image().
   - Various improvements to CMake scripts: e.g. improved warning and error
     messages, added USE_NEON_INSTRUCTIONS option.
   - chol() will use a banded Cholesky algorithm for banded matrices, making it
     much faster in these cases.
   - Changed the timing code to use the C++11 high resolution clock and
     atomics. This makes the timing code a lot more precise.

Non-Backwards Compatible Changes:
   - Changed the random_cropper's set_min_object_size() routine to take min box
     dimensions in the same format as the mmod_options object (i.e. two lengths
     measured in pixels). This should make defining random_cropping strategies
     that are consistent with MMOD settings more straightforward since you can
     simply take the mmod_options settings and give them to the random_cropper
     and it will do the right thing.
   - Changed the mean squared loss layers to return a loss that's the MSE, not
     0.5*MSE. The only thing this effects is the logging messages that print
     during training, which were confusing since the reported loss was half the
     size you might naively expect.
   - Changed the outputs of test_regression_function() and cross_validate_regression_trainer().
     These functions now output 4D rather than 2D vectors.  The new output is:
     mean squared error, correlation, mean absolute error, and standard
     deviation of absolute error.  I also made test_regression_function() take
     a non-const reference to the regression function so that DNN objects can
     be tested.
   - Fixed shape_predictor_trainer padding so it behaves as it used to. In
     dlib 19.7 the padding code was changed and accidentally doubled the size
     of the applied padding in some cases. It's not a huge deal either way, but
     this change reverts back to the previous behavior.

Bug fixes:
   - Fixed toMat() not compiling in some cases.
   - Significantly reduced the compile time of the DNN example programs in
     visual studio.
   - Fixed a few image processing functions that weren't using the generic
     image interface.
   - Fixed a bug in the random_cropper where it might crash due to division by
     0 if small images were given as input.
   - Fixed a bug in how the mmod_options automatically determines detection
     window sizes. It would pick a bad size in some cases.
   - Fixed load_image_dataset()'s skip_empty_images() option. It wasn't
     skipping images that only have ignore boxes when you load into mmod_rect
     objects.
   - Fixed a bug where chinese_whispers(), when called from python, would
     sometimes return a labels array that didn't include labels for all the
     inputs.
   - Fixed a bug in dlib's MS Windows GUI code that was introduced a little
     while back when we switched everything to std::shared_ptr.  This change
     fixes a bug where the program crashes or hangs sometimes during program
     shutdown.
   - Fixed error in TIME_THIS() introduced in dlib 19.7. It was printing
     seconds when it said minutes in the output.
   - Adding missing implementation of tabbed_display::selected_tab.
   - Changed the windows signaler and mutex code to use the C++11 thread
     library instead of the old win32 functions. I did this to work around how
     windows unloads dlls. In particular, during dll unload windows will kill
     all threads, THEN it will destruct global objects. So this can lead to
     problems when a global object that owns threads tries to tell them to
     shutdown, since the threads have already vanished.  The new code mitigates
     some of these problems, in particular, there were some cases where
     unloading dlib's python extension would deadlock.  This should now be
     fixed.
   - Fixed compile time errors when either of these macros were enabled:
     DLIB_STACK_TRACE, DLIB_ISO_CPP_ONLY.

New Features and Improvements:
   - Deep Learning:
      - The CNN+MMOD detector is now a multi-class detector.  In particular,
        the mmod_rect object now has a string label field which you can use to
        label objects, and the loss_mmod_ layer will learn to label objects with
        those labels.  For an example, see: https://www.youtube.com/watch?v=OHbJ7HhbG74
      - CNN+MMOD detectors are now 2.5x faster.  For instance, this example program
        http://dlib.net/dnn_mmod_find_cars_ex.cpp.html now runs at 98fps instead
        of 39fps.  
   - Added a 5 point face landmarking model that is over 10x smaller than the
     68 point model, runs faster, and works with both HOG and CNN generated
     face detections.  It is now the recommended landmarking model to use for 
     face alignment.  render_face_detections() and get_face_chip_details() have been 
     updated to work with both 5 and 68 point models, so the new 5 point model is
     a drop in replacement for the 68 point model.
   - The imglab tool is slightly improved.  It will display box labels with
     higher relative contrast.  You can also now press END or i to ignore boxes
     in imglab. This is useful because it's a much less stressing hand motion
     to hit END that i in most cases.
   - Added overloads of sub_image() that take raw pointers so you can make
     sub_images of anything. 
   - Changed TIME_THIS() to use std::chrono::high_resolution_clock, so now it's
     much higher precision.
   - Exposed Chinese whispers clustering to Python, added face clustering example.

Non-Backwards Compatible Changes:

Bug fixes:
   - Fixed an error in input_rgb_image_pyramid::image_contained_point(). The
     function might erroneously indicate that a point wasn't inside the original
     image when really it was, causing spurious error messages.
   - mmod_options would pick bad window sizes in some corner cases. This has been fixed.
   - Fixed a bug in the extract layer that trigged when a tensor with a
     different number of samples than the tensor used to initialize the network
     was passed through the layer.
   - The loss_per_missed_target parameter of the loss_mmod_ wasn't being used
     exactly right when boxes were auto-ignored.  There weren't any practical
     user facing problems due to this, but it has nevertheless been fixed. 

New Features and Improvements:

Non-Backwards Compatible Changes:

Bug fixes:
   - Fix build error in Visual Studio when CUDA is enabled. 

New Features and Improvements:
   - Deep Learning
      - Added a python wrapper for using the CNN face detector.
      - Added support for cuDNN v6 and v7.
      - Added a simple tool to convert dlib model files to caffe models.  
        See the tools/convert_dlib_nets_to_caffe folder for details.
      - New DNN layers
         - loss_multiclass_log_per_pixel_
         - loss_multiclass_log_per_pixel_weighted_
         - loss_mean_squared_per_pixel_
         - cont_       (transpose convolution, sometimes called "deconvolution")
         - mult_prev_  (like add_prev_ but multiplies instead of adds)
         - extract_    (sort of like caffe's slice layer)
         - upsample_   (upsamples a tensor using bilinear interpolation)
      - Object Detection
         - Upgraded loss_mmod_ to support objects of varying aspect ratio. This
           changes the API for the mmod_options struct slightly.
         - Relaxed the default non-max suppression parameters used by the
           mmod_options object so that users of the deep learning MMOD tool don't
           get spurious errors about impossibly labeled objects during training.
         - Added missing input validation to loss_mmod_.  Specifically, the loss
           layer now checks if the user is giving truth boxes that can't be detected
           because the non-max suppression settings would prevent them from being
           output at the same time. If this happens then we print a warning message
           and set one of the offending boxes to "ignore".  I also changed all
           the input validation errors to warning messages with auto conversion
           to ignore boxes rather than exceptions.
         - Changed the random_cropper's interface so that instead of talking in
           terms of min and max object height, it's now min and max object size.
           This way, if you have objects that are short and wide (i.e. objects where
           the relevant dimension is width rather than height) you will get sensible
           behavior out of the random cropper.
         - Added options to input_rgb_image_pyramid that let the user set
           create_tiled_pyramid()'s padding parameters. Also changed the default
           outer border padding from 0 to 11. This effects even previously trained
           models. So any model that doesn't explicitly set the outer patting to
           something else will have a padding of 11. This should be a more
           reasonable value for most networks.
         - Added process() and process_batch() to add_loss_layer. These routines
           let you easily pass arguments to any optional parameters of a loss
           layer's to_tensor() routine. For instance, it makes it more convenient to
           set loss_mmod_'s adjust_threshold parameter.
      - Added visit_layers_until_tag()
      - Improved how dnn_trainer synchronizes its state to disk.  It now uses
        two files and alternates between them.  This should be more robust in
        the face of random hardware failure during synchronization than the
        previous synchronization method.
      - Made it so you can set the number of output filters for con_ layers at runtime.
      - The way cuDNN work buffers are managed has been improved, leading to
        less GPU RAM usage.  Therefore, users should not need to call
        set_dnn_prefer_smallest_algorithms() anymore.
      - Added operator<< for random_cropper and dnn_trainer to allow 
        easy logging of training parameters.
      - Made concat_ layer a lot faster.
      - Made the dnn_trainer not forget all the previous loss values it knows
        about when it determines that there have been a lot of steps without
        progress and shrinks the learning rate. Instead, it removes only a
        small amount of the oldest values.   The problem with the old way of
        removing all the loss values in the history was that if you set the
        steps without progress threshold to a really high number you would
        often observe that the last few learning rate values were obviously not
        making progress, however, since all the previous loss values were
        forgotten the trainer needed to fully populate its loss history from
        scratch before it would figure this out.  This new style makes the
        trainer not waste time running this excessive optimization of obviously
        useless mini-batches.  I also changed the default
        get_test_iterations_without_progress_threshold() from 200 to 500.  Now
        that we have a better history management of loss values in the trainer
        it's much more sensible to have a larger value here.
   - Dlib's simd classes will now use ARM NEON instructions.  This makes the
     HOG based object detector faster on mobile devices running ARM processors.
   - Added last_modified() method to dlib::file.  Also, added
     select_oldest_file() and select_newest_file().
   - Added solve_qp_box_constrained_blockdiag()
   - Added an overload of mat() that takes a row stride value.
   - Added cmake scripts and some related tooling that makes it easy to call
     C++ code from java.  See dlib/java/ folder.  
   - MATLAB MEX wrapper API
      - Made the mex wrapper deal with cell arrays that have null elements.
      - Made ctrl+c detection in a mex file work more reliably in newer versions of matlab.
   - Added set_rect_area()
   - Gave test_object_detection_function() an option to set how ignore box
     overlap is tested.
   - Added serialization support for the running_stats_decayed object.
   - Additions to imglab
      - Added --sort and also the ability to propagate boxes from one image to
        the next using dlib::correlation_tracker.
      - Made it so you can remove images by pressing alt+d. 
      - Made is so pressing e in imglab toggles between views of the image
        where the histogram is equalized or unmodified. This way, if you are
        looking at particularly dark or badly contrasted images you can toggle
        this mode and maybe get a better view of what you are labeling.
   - Made the attribute_list of the xml parser a little more friendly by
     allowing you to ask for attributes that don't exist and get a defined
     behavior (an exception being thrown) rather than it being a contract
     violation.

Non-Backwards Compatible Changes:
   - DNN solver objects are now required to declare operator<<.
   - Broke backwards compatibility with previous dnn_trainer serialization
     format.  The network serialization format has not changed however.  So old
     model files will still load properly.
   - Changed random_cropper interface.
   - Changed the XML format output by net_to_xml(). Specifically, the XML tag
     for affine layers was changed to use the same conventions as other layers
     that support convolutional vs fully connected modes.
   - Dlib's smart pointers have been deprecated and all of dlib's code has been
     changed to use the std:: version of these smart pointers.  The old dlib
     smart pointers are still present, allowing users to explicitly include
     them if needed, but users should migrate to the C++11 standard version of
     these tools. 
   - Changed the functions that transform between input tensor coordinates and
     output tensor coordinates to use dpoint instead of point. This way, we can
     obtain sub-pixel coordinates if we need them.
   - Upgraded loss_mmod_ to support objects of varying aspect ratio. This
     changes the API for the mmod_options struct slightly.

Bug fixes:
   - Made resize_image() and functions that use it like the pyramid objects
     produce better results when run on float and double images. There was
     needless rounding to integers happening in the bilinear interpolation. Now
     if you work with a float image the entire process will run without integer
     rounding.
   - Made the input_tensor_to_output_tensor() and output_tensor_to_input_tensor() 
     coordinate mappings work on networks that contain skip layers.
   - The input_rgb_image_sized is supposed to be convertible to
     input_rgb_image, which it was in all ways except you couldn't deserialize
     directly like you would expect. This has now been fixed.
   - There was a bug in the concat_ layer's backward() method. It was assigning
     the gradient to previous layers instead of adding the gradient, as required
     by the layer interface specification.  Probably no-one has been impacted
     by this bug, but it's still a bug and has been fixed.
   - Changed the random_cropper so that it samples background patches uniformly
     across scales regardless of the input image size. Previously, if you gave
     really large images or really small images it had a bias towards giving only
     large patches or small patches respectively.
   - Fixed name lookup problem for calls to serialize() on network objects.
   - Fixed double delete in tokenizer_kernel_1.
   - Fixed error in pyramid_down<2> that caused the output image to be a
     little funny looking in some cases.
   - Fixed the visit_layers_backwards() and visit_layers_backwards_range()
     routines so they visit layers in the correct order.
   - Made build scripts work on a wider range of platforms and configurations.
   - Worked around global timer cleanup issues that occur on windows when dlib
     is used in a dll in some situations.
   - Fixed various compiler errors in obscure environments.

New Features:

Non-Backwards Compatible Changes:
   - CMake 2.8.12 is now required to build dlib (but only if you use CMake).  

Bug fixes:
   - Fixed a slow memory leak that could occur when using cuDNN.

Other:

New Features:
   - Deep Learning
      - Added a state-of-the-art face recognition tool (99.38% accuracy on the
        LFW benchmark) with C++ and Python example programs.
      - Added these new loss layer types: loss_metric_, loss_mean_squared_, and
        loss_mean_squared_multioutput_.
      - Added the l2normalize_ computational layer.
      - Added test_one_step() to the dnn_trainer. This allows you to do
        automatic early stopping based on observing the loss on held out data.
      - Made the dnn_trainer automatically reload from the last good state if a
        loss of NaN is encountered.
      - Made alias_tensor usable when it is const.
   - Dlib's simd classes will now use PowerPC VSX instructions.  This makes the
     HOG based object detector faster on PowerPC machines.
   - Added compute_roc_curve()
   - Added find_gap_between_convex_hulls()
   - Added serialization support for std::array.
   - Added running_scalar_covariance_decayed object
   - Added running_stats_decayed object
   - Added min_pointwise() and max_pointwise().
   - Added a 1D clustering routine: segment_number_line().
   - Added Intel MKL FFT bindings.
   - Added matlab_object to the mex wrapper. Now you can have parameters that
     are arbitrary matlab objects.
   - Added support for loading of RGBA JPEG images

Non-Backwards Compatible Changes:
   - Changed the loss layer interface to use two typedefs, output_label_type
     and training_label_type instead of a single label_type. This way, the label
     type used for training can be distinct from the type output by the network.
     This change breaks backwards compatibility with the previous API.

Bug fixes:
   - Fixed compiler warnings and errors on newer compilers.
   - Fixed a bug in the repeat layer that caused it to throw exceptions in some
     cases.
   - Fixed matlab crashing if an error message from a mex file used the %
     character, since that is interpreted by matlab as part of an eventual
     printf() code.
   - Fixed compile time error in random_subset_selector::swap()
   - Fixed missing implementation of map_input_to_output() and
     map_output_to_input() in the concat_ layer.
   - Made the dnn_trainer's detection and backtracking from situations with
     increasing loss more robust. Now it will never get into a situation where it
     backtracks over and over. Instead, it will only backtrack a few times in a
     row before just letting SGD run unimpeded.

Other:
   - Usability improvements to DNN API.
   - Improved C++11 detection, especially on OS X.
   - Made dlib::thread_pool use std::thread and join on the threads in
     thread_pool's destructor. The previous implementation used dlib's global
     thread pooling to allocate threads to dlib::thread_pool, however, this
     sometimes caused annoying behavior when used as part of a MATLAB mex file,
     very occasionally leading to matlab crashes when mex files were unloaded.
     This also means that dlib::thread_pool construction is a little bit slower
     than it used to be.

New Features:
   - Updates to the deep learning API:
      - Added tools for making convolutional neural network based object detectors.  See
        dnn_mmod_ex.cpp example program.
      - Added annotation() to tensor so you can associate any object you want with a tensor.
      - Made layer_details() part of the SUBNET interface so that user defined layer
        details objects can access each other. Also added the input_layer() global function
        for accessing the input layer specifically.
      - alias_tensor can now create aliases of const tensors.
      - Added set_all_bn_running_stats_window_sizes().
      - Added visit_layers_backwards(), visit_layers_backwards_range(), and
        visit_layers_range().
      - Computational layers can now optionally define map_input_to_output() and
        map_output_to_input() member functions.  If all layers of a network provide these
        functions then the new global functions input_tensor_to_output_tensor() and
        output_tensor_to_input_tensor() can be used to map between the network's input and
        output tensor coordinates.  This is important for fully convolutional object
        detectors since they need to map between the image space and final feature space.
        These new functions are important for tools like the new MMOD detector.
      - Added input_rgb_image_pyramid.
   - Image Processing
      - The imglab command line tool has these new options: --min-object-size, --rmempty,
        --rmlabel, --rm-if-overlaps, and --sort-num-objects.  I also changed the behavior of
        --split so that it simply partitions the data and is an invertible operation.
      - Added mmod_rect
      - Added an overload of load_image_dataset() that outputs directly to mmod_rect
        instead of rectangle.
      - Added image_dataset_file::shrink_big_images(). So now load_image_dataset() can load
        a dataset of high resolution files at a user requested lower resolution.
      - Added box_intersection_over_union().
      - Added create_tiled_pyramid(), image_to_tiled_pyramid(), and tiled_pyramid_to_image().
      - Added random_cropper
   - Upgraded dlib's mex wrapper tooling to enable easy binding of C++ classes to MATLAB
     objects. 
   - Added nearest_rect()
   - Added find_upper_quantile()
   - Added count_steps_without_decrease_robust().
   - Added get_double_in_range() to dlib::rand.

Non-Backwards Compatible Changes:
   - C++11 is now required to use dlib.  
   - Changed pinv() so it interprets its tol argument relative to the largest singular
     value of the input matrix rather than as an absolute tolerance.  This should generally
     improve results, but could change the output in some cases.
   - Renamed the class members of test_box_overlap so they are less confusing.
   - Updates to the deep learning API:
      - Changed the DNN API so that sample_expansion_factor is a runtime variable rather
        than a compile time constant. This also removes it from the input layer interface
        since the DNN core now infers its value at runtime. Therefore, users that define their
        own input layers don't need to specify it anymore.
      - Changed DEFAULT_BATCH_NORM_EPS from 1e-5 to 1e-4.
      - Changed the default batch normalization running stats window from 1000 to 100.

Bug fixes:
   - Made the relational operators constexpr so they don't accidentally cause compilation
     errors when they get pulled into the scope of template metaprogramming expressions.
   - Fixed all/source.cpp not compiling in some instances.
   - CMake scripts now do a better job detecting things like C++11 support, the presence of
     CUDA, and other system specific details that could cause the build to fail if not
     properly configured.
   - Fixed a bug in imglab's --cluster option where it would output xml files with empty
     entries if the input xml file contained unannotated images.
   - Fixed imglab's --cluster option not working with relative paths.

Other:
   - Made the thread local variables that hold the cudnn and cublas context objects not
     destruct and recreate themselves when you switch devices.  Instead, they keep a table
     of context objects, for each thread and device, reusing as necessary. This prevents
     churn in the context objects when you are switching back and forth between devices
     inside a single thread, making things run more efficiently for some CUDA based
     workflows.
   - Made the message argument of the DLIB_ASSERT and DLIB_CASSERT macros optional.
   - Made thread_pool and parallel_for propagate exceptions from task threads to calling
     code rather than killing the application if a task thread throws.
   - Changed imglab --resample so that it never changes the aspect ratio of an image.
   - Made the check in dnn_trainer for convergence more robust. Previously, if we
     encountered a bad mini-batch that made the loss value suddenly jump up by a larger than
     normal value it could make the trainer think we converged. Now the test is robust to
     transient spikes in loss value.  Additionally, the dnn_trainer will now check if the
     loss has been increasing before it saves the state to disk. If it detects that the loss
     has been going up then instead of saving to disk it recalls the previously good state.
     This way, if we hit a really bad mini-batch during training which negatively effects
     the model in a significant way, the dnn_trainer will automatically revert back to an
     earlier good state.

New Features:
   - Support for cuDNN 5.1
   - dlib::async() and dlib::default_thread_pool().
   - rectangle_transform
   - imglab tool: added --resample, --ignore, --files, and --extract-chips
     command line options.  Also added convert_imglab_paths_to_relative and
     copy_imglab_dataset scripts.
   - Evgeniy Fominov made the shape_predictor trainer multi-threaded and faster.
   - sutr90 contributed support for the CIELab color space.  See the new lab_pixel.

Non-Backwards Compatible Changes:
   - All the cmake utility scripts were moved to dlib/cmake_utils.  
   - Code that #includes the shape_predictor can now only be compiled with
     compilers that support C++11 lambda functions.

Bug fixes:
   - Made CMake scripts work in a wider range of environments. 
   - Fixed compile time errors on various platforms.
   - Fixed bad multi-threading support in the MATLAB mex wrapper.
   - Fixed bug in cuDNN binding that could sometimes cause NaN outputs.
   - Fixed bad convergence testing in DNN tooling for very small datasets.

Other:

New Features:
   - A deep learning toolkit using CPU and/or GPU hardware.  Some major elements
     of this are:
      - Clean and fully documented C++11 API
      - Clean tutorials: see dnn_introduction_ex.cpp and dnn_introduction2_ex.cpp
      - Uses cuDNN v5.0
      - Multi-GPU support
      - Automatic learning rate adjustment
      - A pretrained 1000 class Imagenet classifier (see dnn_imagenet_ex.cpp)
   - Optimization Tools
      - Added find_optimal_parameters()
      - Added elastic_net class
      - Added the option to use the elastic net regularizer to the OCA solver.
      - Added an option to solve the L2-loss version of the SVM objective function to svm_c_linear_dcd_trainer.
      - Added solve_qp_box_constrained()
   - Image Processing
      - Added random_color_transform, disturb_colors(), and apply_random_color_offset().
      - load_image() now supports loading GIF files.
   - Many improvements to the MATLAB binding API  
      - Automatically link to MATLAB's Intel MKL when used on linux.
      - struct support
      - mex functions can have up to 20 arguments instead of 10.
      - In place operation.  Made column major matrices directly wrap MATLAB
        matrix objects when used inside mex files.  This way, if you use
        matrix_colmajor or fmatrix_colmajor in a mex file it will not do any
        unnecessary copying or transposing.
      - Catch ctrl+c presses in MATLAB console.  Allowing early termination of mex functions.
      - When used inside mex files, DLIB_ASSERTS won't kill the MATLAB process,
        just throw an exception.
      - Made cerr print in MATLAB as a red warning message.
   - load_mnist_dataset()
   - Added a constructor for seeding rand with a time_t.
   - Added subm_clipped()
   - Added unserialize.
   - Added running_gradient

Non-Backwards Compatible Changes:
   - Everything in dlib/matlab/call_matlab.h is now in the dlib namespace.
   - DLIB_TEST() and DLIB_TEST_MSG() macros now require you to terminate them with a ;

Bug fixes:
   - Fixed bug in 10 argument version of call_matlab() and also cleaned up a few
     minor things.
   - setup.py and CMake scripts work in a few more contexts.
   - Fixed compiler errors in visual studio 2015.
   - Fixed a bug in gaussian_blur() that caused messed up outputs when big
     sigma values were used on some pixel types.
   - Fixed minor bugs in join_rows() and join_cols(). They didn't work when one
     of the matrices was empty.

Other:
   - Made CMake scripts uniformly require CMake version 2.8.4.
   - Faster fHOG feature extraction / face detection
   - CMake scripts now enable C++11 by default
   - Gave array2d and matrix move constructors and move assignment operators.  Matrix
     can also now be created from initializer lists.

New Features:
   - Added the set_ptrm() routine for assigning dlib::matrix objects to arbitrary
     memory blocks.
      
Non-Backwards Compatible Changes:

Bug fixes:
   - Fixed a bug that caused cmake to not provide the correct preprocessor
     definitions until cmake was run twice. This was causing some projects to
     not build properly.

Other:
   - Improvements to build system:
      - Ehsan Azarnasab contributed a setup.py so the dlib Python API can be
        installed via the usual 'python setup.py install' command. 
      - Séverin Lemaignan upgraded dlib's CMake scripts so they include an 
        install target.  Now dlib can be installed system wide by executing 
        'cmake PATH_TO_DLIB; make install'.  This also includes installing the
        appropriate scripts for CMake's find_package(dlib) to work.

New Features:
   - More clustering tools:
      - Added bottom_up_cluster() and find_clusters_using_angular_kmeans()
        routines.
      - Added a --cluster option to the imglab tool.  This lets you cluster
        objects into groups of similar appearance/pose.
   - Improved the shape_predictor.  In particular, it can now be learned from
     datasets where some landmarks are missing.  The shape_predictor also now
     outputs a sparse feature vector that encodes which leafs are used on each
     tree to make a prediction.
      
Non-Backwards Compatible Changes:
   - extract_highdim_face_lbp_descriptors() produces slightly different output.

Bug fixes:
   - Fixed a minor bug in extract_highdim_face_lbp_descriptors() which was
     pointed out by Yan Xu. One of the face locations was mistakenly used twice
     while another was skipped. This change breaks backwards compatibility with
     the previous feature extraction output but should slightly improve
     accuracy of classifiers trained using these features.
   - Fixed jet() and heatmap() so they work on empty images.
   - The SQLite transaction object did not function correctly when compiled 
     in a C++11 program.  Since its destructor can throw, an exception
     specification needed to be added indicating that this was possible since
     destructors are now noexcept by default in C++11.
   - Fixed a bug pointed out by Ernesto Tapia that could cause matrix
     expressions that involve sub matrix views (e.g. colm) to produce the wrong
     results when the BLAS bindings were enabled.
   - Added an if to avoid a possible division by zero inside spectral_cluster().
   - Fixed a bug in parse_xml(). It failed to check if the given input stream
     was valid before trying to parse it.

Other:

New Features:
   - Added a linear model predictive control solver.  See the mpc_ex.cpp example
     program for details.
   - Thanks to Patrick Snape, the correlation_tracker can now be used from Python.
      
Non-Backwards Compatible Changes:
   - The camera_transform's second operator() method now takes 3 arguments
     instead of 2.  This is to allow it to output the z distance in addition to
     scale.

Bug fixes:
   - Fixed a bug in the eigenvalue_decomposition which could occur when a
     symmetric matrix was used along with the LAPACK bindings.
   - Fixed a bug where the last column of data in a file wasn't loaded on some
     OS X machines when load_libsvm_formatted_data() was called.

Other:
   - Added a hard iteration limit to a number of the SVM solvers.
   - Adrian Rosebrock graciously setup an OS X machine for dlib testing, which
     resulted in improved CMake python scripts on OS X machines.
   - Improved the way overlapping points are rendered by the perspective_window.

New Features:
   - Added a number of tools for working with 3D data:
      - Added the perspective_window which is a tool for displaying 3D point clouds.
      - Added camera_transform.  It performs the 3D to 2D mapping needed to visualize 3D
        data.
      - Added point_transform_affine3d as well as functions for creating such transforms:
        rotate_around_x(), rotate_around_y(), rotate_around_z(), and translate_point().
   - Added draw_solid_circle() for drawing on images.
   - Added get_best_hough_point() to the hough_transform.
   - Thanks to Jack Culpepper, the python API for object detection now outputs detection
     confidences.  
   - Added lspi, an implementation of the least-squares policy iteration algorithm.
      
Non-Backwards Compatible Changes:
   - The shape_predictor and shape_predictor_trainer had a non-optimal behavior when used
     with objects that have non-square bounding boxes. This has been fixed but will cause
     models that were trained with the previous version of dlib to not work as accurately if
     they used non-square boxes. So you might have to retrain your models when updating dlib.

Bug fixes:
   - Fixed a bug which prevented add_image_rotations() from compiling.

Other:
   - The imglab tool now allows the user to click and drag annotations around by holding
     shift and right clicking.

New Features:
   - Added spectral_cluster()
   - Added sub_image() and sub_image_proxy
   - Added set_all_logging_headers()

Non-Backwards Compatible Changes:

Bug fixes:
   - Fixed a bug that caused the correlation_tracker to erroneously trigger an assert when
     run in debug mode.

Other:
   - Improved the usability of the new drectanle object.
   - Optimized extract_fhog_features() for the case where cell_size==1. This makes it about
     4x faster in that case.
   - Made it so you can compose point transform objects via operator *.

New Features:
   - Added the correlation_tracker object
   - Added the option to force the last weight to 1 to structural_assignment_trainer.
   - Added max_point_interpolated()
   - Added the drectangle object
   - New Python Tools:
      - Patrick Snape contributed a Python binding for the face landmarking tool and 
        the general purpose shape prediction/training tools.
      - Vinh Khuc contributed a Python binding for find_candidate_object_locations(),
        dlib's implementation of the selective search object location proposal method.

Non-Backwards Compatible Changes:

Bug fixes:
   - Fixed a bug in extract_image_chips() and get_mapping_to_chip() that caused
     incorrect outputs when the requested chip stretched the image unevenly
     vertically or horizontally.
   - Made CMake check that libpng and libjpeg actually contain the link symbols
     they are supposed to since, on some systems, these libraries aren't
     installed correctly and will cause linker errors if used.
   - Fixed assign_border_pixels(img, rect) so that it correctly zeros an image
     when an empty rectangle is supplied. Previously, it did nothing to the
     image in this case.
   - Fixed compute_lda_transform() so it works properly when the class
     covariance matrices are singular even after performing PCA.
   - Fixed a bug in find_similarity_transform(). When given just two points as
     inputs it would sometimes produce a reflection rather than a similarity
     transform.
   - Disabled all bindings to FFTW because FFTW isn't threadsafe.

Other:
   - Added an example program for dlib's SQLite API and made a few minor
     usability improvements to the API as well.

New Features:
   - Upgraded fft() and ifft() to support 2D matrices.
   - Added hough_transform
   - Added skeleton() for finding the skeletonization of a binary image.
   - Added distance_to_line(), clip_line_to_rectangle(), min_point(), and max_point().
   - Added a simple API for calling C++ from MATLAB.

Non-Backwards Compatible Changes:

Bug fixes:
   - Fixed a compile time error that could happen when calling fft() for
     certain input types.
   - Fixed a compile time error that prevented auto_threshold_image() from
     being used.
   - Fixed name clashes with new version of Boost. 
   - Changed Python pickling code so it works with Python 3.
   - Fixed CMake compile time error related to finding fftw.

Other:
   - Made extract_image_chips() much faster when extracting unscaled image chips.

New Features:
   - Added save_jpeg()
   - Added the option to use an identity matrix prior to vector_normalizer_frobmetric.
   - Made the extract_image_chips() routine more flexible, in particular: Added
     get_mapping_to_chip(), get_face_chip_details(), map_det_to_chip(), and also
     upgraded chip_details so you can specify a chip extraction by a bunch of
     point correspondences between the chip and the original image.
   - Added a set of local-binary-pattern based feature extractors:
     make_uniform_lbp_image(), extract_histogram_descriptors(),
     extract_uniform_lbp_descriptors(), and extract_highdim_face_lbp_descriptors() 
   - Added compute_lda_transform() 
   - Added equal_error_rate()
   - Added cast_to() to the type_safe_union. This allows you to get the
     contents of a const type_safe_union.

Non-Backwards Compatible Changes:

Bug fixes:
   - Changed noncopyable.h to avoid a name clash with boost 1.56
   - On some platforms hostname_to_ip() would erroneously return 0.0.0.0. This
     has been fixed.

Other:

New Features:
   - Added find_similarity_transform()
   - Added the ability to upgrade a auto_mutex_readonly from a readonly lock to a write
     lock. 
   - Added an implementation of the paper "One Millisecond Face Alignment with an Ensemble
     of Regression Trees" by Vahid Kazemi and Josephine Sullivan which appeared in this
     year's CVPR conference.  Therefore, dlib now includes tools for learning shape models
     and also comes with a state-of-the-art face landmark locator.  See the
     face_landmark_detection_ex.cpp and train_shape_predictor_ex.cpp example programs for
     an introduction.

Non-Backwards Compatible Changes:
   - Made the interface to all the image processing routines more generic.  In particular, 
     it is now easier to use arbitrary image types with dlib.  The new generic image
     interface is defined in dlib/image_processing/generic_image.h and simply consists of
     seven user defined global functions and a traits template.  Any user code that was
     using array2d objects to represent images will still work.  However, if you had been
     using your own custom image object you will need to provide implementations of the
     seven functions.  Instructions for how to do this are in
     dlib/image_processing/generic_image.h.

Bug fixes:
   - Changed the murmur hash implementation to avoid any possibility of strict aliasing
     violations in user code, even when things get inlined in unfavorable ways.
   - Fixed a color space handling bug in resize_image() that caused bad looking outputs in
     some cases.
   - If "cmake" was a substring of the full path to your source code folder then the cmake
     scripts would fail. This has been fixed.
   - Fixed a compile time error that could occur when using find_max_single_variable().

Other:
   - load_image() now uses the internal file header information to detect the
     image format rather than looking at the file extension.
   - Renamed unit test program to dtest avoid warnings from CMake.
   - cross_validate_trainer() and cross_validate_trainer_threaded() no loner make copies
     of the training data.  This significantly reduces their RAM usage for large datasets.
   - Changed the serialization code for C-strings so that they don't save the null
     terminator byte. This makes their serialization format the same as the format for
     std::string.  The code should still be able to read all previously serialized data
     correctly, so the change is backwards compatible with previous versions of dlib. 
   - Changed the evaluate_detectors() routine so that it applies non-max suppression to
     each detector individually. This way one detector doesn't stomp on the output of
     another detector.
   - Made the version of draw_line() that draws onto a regular image use alpha blending
     for drawing diagonal lines.

New Features:

Non-Backwards Compatible Changes:

Bug fixes:
   - The new simplified serialization API that works like serialize("filename")<<object 
     was not opening files in binary mode and therefore didn't work properly on Windows.  
     This has been fixed.

Other:

New Features:
   - Added the ability to set a previously trained function as a prior to the
     svm_multiclass_linear_trainer, svm_c_linear_trainer, and svm_rank_trainer
     objects.
   - Added a user settable loss to the structural_assignment_trainer and
     structural_track_association_trainer objects.  
   - Added evaluate_detectors(), a function for efficiently running multiple fHOG
     based object detectors.
   - Added the new split_on_first() and split_on_last() string manipulation functions.
   - Added locally_change_current_dir, a RAII tool for switching between directories.
   - You can now make a 1x1 matrix containing a single value by calling mat() on a single
     scalar value.
   - The point transform functions and frobmetric_training_sample are now serializable.
   - Added a simplified operator << and >> based syntax for serializing to and
     from files.  So now you can serialize to a file using a syntax of:
       serialize("myfile.dat") << myobject << another_object;
     and then load those objects from disk via:
       deserialize("myfile.dat") >> myobject >> another_object;
     An arbitrary number of objects can be serialized or deserialized by
     chaining the << and >> operators.

Non-Backwards Compatible Changes:

Bug fixes:
   - Fixed a bug pointed out by Daniel Girardeau-Montaut. The covariance()
     function didn't work on non-double valued matrices.
   - Fixed a bug in the backtracking_line_search() function pointed out by
     Ping-Chang Shih. The function ignored the max_iter parameter.
   - Fixed a compiler error encountered when using clang 3.4 on Mac OS X 10.9.
     Thanks to Martin Fergie for reporting this problem.
   - Fixed a potential divide by zero in draw_fhog()

Other:
   - Added an example program showing how to set a custom logger output hook.
   - Made linear decision_functions which use sparse vectors much faster.

New Features:
   - Added a Python API for working with fHOG based object detectors.  See the
     new python example programs train_object_detector.py and face_detector.py for
     more details.
   - Added the ability to use a user supplied fHOG style feature extractor with 
     the scan_fhog_pyramid object.  So now you can define your own version of HOG
     for use with these tools.
   - The oca solver now supports taking a user supplied prior vector. That is,
     it lets you use a regularizer like ||w-prior||^2 instead of the usual
     ||w||^2 regularizer.
   - Added the structural_track_association_trainer object.  It is a structural
     SVM tool for creating multi-target tracking algorithms.  See the
     learning_to_track_ex.cpp example program for an introduction.
   - Added the following minor utility functions: nearest_center(),
     add_image_rotations(), set_aspect_ratio(), and tile_images().

Non-Backwards Compatible Changes:
   - Refactored the load_image_dataset() routines so they are easier to use and
     more flexible. This introduces a slight backwards incompatibility in that
     the version that loads full_object_detection objects now returns an ignore
     rectangle set instead of a parts name list.  Other than that the changes
     are backwards compatible with previous versions of dlib.
   - Added a bias term to the assignment_function's model so the user doesn't
     need to remember, or even understand, that they should add it themselves.
     However, this change breaks backwards compatibility with the previous
     serialization format for assignment_function objects.

Bug fixes:
   - Fixed a number of compile time errors that could occur in rare cases.
   - The stopping condition for the svr_linear_trainer was too tight, causing it
     to take an excessive amount of time to converge in some cases. 
   - Disabled use of XIM for X11 windowing since it makes programs hang on some
     systems. However, this means the wide character input methods won't work on
     X11 systems anymore.
   - Fixed a bug in randomize_samples() which caused the outputs to be not as
     random as they should be. 
   - Fixed dlib's CMakeLists.txt file so that the "use FFTW" option actually
     causes the build to use FFTW.
   - Fixed a compile time error that triggered when trying to link with FFTW.
   - mat() did not work correctly when used with std::vector<bool> objects.
     This has been fixed.

Other:

New Features:
   - Object Detection Tools:
      - Added scan_fhog_pyramid, a tool for creating Histogram of Oriented Gradient (HOG)
        based object detectors. 
      - Added get_frontal_face_detector(), a HOG based face detector.  
      - Added an option to include "ignore/don't care" truth boxes to the
        structural_object_detection_trainer.  This allows a user to tell the trainer that
        they don't care if certain objects are detected or not detected.  
   - Image Processing Tools:
      - Added extract_image_chips()
      - Added a version of draw_rectangle() for drawing on images.
      - The spatial filtering routines now support even sized filters.
      - Added flip_image_dataset_left_right(), upsample_image_dataset(), and
        rotate_image_dataset().
   - Machine Learning Tools:
      - Added a nuclear norm regularization option to the structural SVM solver.
      - Added the option to learn only non-negative weights to the
        svm_multiclass_linear_trainer.
   - Speed Improvements:
      - The svm_multiclass_linear_trainer, one_vs_one_trainer, and one_vs_all_trainer
        objects are now multithreaded.  This also means you have to #include
        dlib/svm_threaded.h instead of dlib/svm.h to use these tools. 
      - A number of image processing tools can now optionally use SSE and AVX instructions
        and are therefore considerably faster.  In particular, the following tools have been
        accelerated: extract_fhog_features, resize_image, pyramid_down, pyramid_up,
        spatially_filter_image_separable, and spatially_filter_image.
   - Added an inv() routine that inverts point transformation functions.
   - Added a sign() routine for matrix objects.

Non-Backwards Compatible Changes:
   - The spatial image filtering functions have the following changes:
      - They no longer zero the image borders when you set the add_to parameter to true.  
      - The spatially_filter_image_separable_down() routine now only allows grayscale
        output images.
   - Changed the default parameters of the test_box_overlap object. Now it defaults to
     using exactly the PASCAL VOC match criterion.
   - To use the svm_multiclass_linear_trainer, one_vs_one_trainer, or one_vs_all_trainer
     objects you now have to #include dlib/svm_threaded.h instead of dlib/svm.h. 
   - pyramid_up() no longer has a levels option.

Bug fixes:
   - Fixed a compile time bug that could occur when wide character strings were
     serialized.
   - Fixed a compile time bug occurring in gcc 4.7.1 on SUSE Linux.  Thanks to Volker
     Härtel for finding this.
   - Fixed compile time errors that occurred when using gcc on cygwin.
   - Fixed a compile time bug that could occur when serializing mlp objects.
   - Fixed a bug in the bigint object that caused division to sometimes produce incorrect
     results.
   - Fixed a bug which sometimes caused load_image_dataset() to erroneously report that
     the dataset file could not be found.
   - Fixed a bug in the structural_object_detection_trainer that caused it to erroneously
     throw a impossible_labeling_error exception in certain rare cases.
   - Updated find_max_factor_graph_nmplp() to use the improved version of the algorithm
     from the 2011 paper Introduction to dual decomposition for inference by David Sontag,
     Amir Globerson, and Tommi Jaakkola.  The original algorithm presented in their 2008
     paper had an error that negatively affected its convergence.  Thanks to James Gunning
     for pointing this out.

Other:
   - Fixed many compiler warnings in gcc 4.8.
   - Made many of the mat() converters bind the resulting matrix expressions into BLAS
     functions.
   - libpng and libjpeg are now included in the dlib/external folder to enable easy static
     linking to these libraries on platforms that typically don't have them (e.g. Windows).
     Moreover, dlib's cmake files will automatically perform this static linking when no
     copy of these libraries is found on the system. 

New Features:
   - Added routines for performing BFGS and L-BFGS optimization with box constraints.  
     See the new find_min_box_constrained() and find_max_box_constrained() routines.
   - Added vector_normalizer_frobmetric.  This is a tool for learning a
     Mahalanobis distance metric.
   - The user can now set different loss values for false alarming vs. getting a
     correct detection when using the structural_sequence_segmentation_trainer.
   - Added an overload of clamp() that lets you use matrix valued lower/upper bounds.
   - New image processing tools:
      - Added the scan_image_custom object, split_array(), and add_image_left_right_flips().
      - Added extract_fhog_features(), this is a function for computing
        Felzenszwalb's 31 channel HOG image representation.  

Non-Backwards Compatible Changes:
   - Refactored the image pyramid code. Now there is just one templated object called
     pyramid_down and you give it the downsampling amount as a template argument.  To make
     old code work with this change use the following substitutions:
       change pyramid_down to pyramid_down<2>
       change pyramid_down_3_2 to pyramid_down<3>
       change pyramid_down_4_3 to pyramid_down<4>
       change pyramid_down_5_4 to pyramid_down<5>

Bug fixes:

Other:
   - Made the structural SVM solver slightly faster.
   - Moved the python C++ utility headers from tools/python/src into dlib/python.
   - The PNG loader is now able to load grayscale images with an alpha channel.
   - Removed checks that prevented users from using references to functions with the
     optimization code and forced the use of function pointers. This was to avoid
     triggering a bug in gcc 4.0.  Since that compiler is no longer officially supported
     by dlib I've removed these checks to increase usability.
   - Made resize_image() use bilinear interpolation by default and also added a special
     version of it that is optimized for this case.
   - Dlib's cmake files will now automatically link to the Intel MKL on MS Windows
     platforms if the MKL is installed.

New Features:
   - Added Python interfaces to dlib's structural support vector machine solver and
     Hungarian algorithm implementation.
   - Added running_cross_covariance
   - Added order_by_descending_distance()
   - Added is_finite()
   - Added the csv IO manipulator that lets you print a matrix in comma separated value
     format.

Non-Backwards Compatible Changes:
   - Changed the object detector testing functions to output average precision instead of
     mean average precision.
   - Added an option to weight the features from a hashed_feature_image relative to the
     number of times they occur in an image.  I also made it the default behavior to use
     this relative weighting and changed the serialization format to accommodate this.

Bug fixes:
   - Fixed typo in learn_platt_scaling(). The method wasn't using the exact prior
     suggested by Platt's paper.
   - Fixed a bug in running_scalar_covariance that caused the covariance() and
     correlation() methods to output the wrong answer if the covariance was negative.

Other:
   - Gave the image_window the ability to tie the mouse and keyboard events together such
     that it is possible for a user to listen for both simultaneously.
   - A number of changes were made to the structural_svm_problem's code which make it 
     significantly faster in some cases.
   - Added Steven Van Ingelgem's patch to the HTTP server which makes operations on HTTP
     headers case-insensitive.

New Features:
   - Machine Learning:
      - Added the svr_linear_trainer, a tool for solving large scale support vector 
        regression problems.
      - Added a tool for working with BIO and BILOU style sequence taggers/segmenters.  
        This is the new sequence_segmenter object and its associated 
        structural_sequence_segmentation_trainer object.
      - Added a python interface to some of the machine learning tools.  These
        include the svm_c_trainer, svm_c_linear_trainer, svm_rank_trainer, and
        structural_sequence_segmentation_trainer objects as well as the cca()
        routine.  
   - Added point_transform_projective and find_projective_transform().
   - Added a function for numerically integrating arbitrary functions, this is the 
     new integrate_function_adapt_simpson() routine which was contributed by 
     Steve Taylor
   - Added jet(), a routine for coloring images with the jet color scheme.

Non-Backwards Compatible Changes:

Bug fixes:
   - Fixed a bug in hysteresis_threshold() that caused it to produce incorrect
     outputs in some cases.
   - Fixed a segmentation fault in the eigenvalue_decomposition object which
     could occur when NaN valued inputs were given.

Other:
   - Made image saving routines work on matrix objects in addition to array2d objects.
   - The machine learning page now contains a flow chart to help new users
     select a machine learning tool appropriate for their task.

New Features:
   - Object Detection Tools:
      - Added another image scanning tool similar to scan_image_pyramid.  This
        is the new scan_image_boxes object.  It allows a user to easily specify
        an arbitrary set of object boxes which should be checked by an object
        detector rather than scanning a fixed sized window over the image as is
        done by the scan_image_pyramid tool.  This allows more flexible scanning 
        strategies.  For example, it is now possible to use the selective search 
        method implemented by the new find_candidate_object_locations() routine.  
      - Added the binned_vector_feature_image.
      - Upgraded the object_detector so that you can use the adjust_threshold
        argument for all versions of the operator() method.
      - Added remove_unobtainable_rectangles()
   - Hashing Tools:
      - Added a set of new locality sensitive hashing functions meant for use
        with larger vectors and higher bit sizes than the current LSH tools.
        These are the new hash_similar_angles_xxx objects.
      - Added find_k_nearest_neighbors_lsh() and hash_samples()
      - Added create_max_margin_projection_hash()
   - New Matrix Routines: linpiece(), fft(), and ifft()
   - Added numeric constants and additional statistics to the running_stats
     object.  This code was contributed by Steve Taylor.
   - Added the image_window::get_next_keypress() routine.  This tool allows a
     user to easily find out which keyboard key the user is pressing.

Non-Backwards Compatible Changes:
   - Changed the object_detector interface slightly.  In particular, it no
     longer adds the adjust_threshold argument to the output scores.    
   - The object detector testing functions now output mean average precision in
     addition to precision and recall.
   - Changed how dlib does serialization in a number of ways:
      - The running_stats and scan_image_pyramid objects have had their
        serialization format changed in a way that breaks backwards
        compatibility.  This means serialized versions of these objects can't be
        loaded by the new version of dlib.
      - Changed the format dlib uses when it serializes floating point values.
        Previously, we used an ASCII based format.  Dlib now uses a much more
        efficient binary format. However, the deserialization routines have
        been made backwards compatible with the previous format. So dlib can
        still deserialize older data but older software won't be able to read
        the new format.
      - Changed the serialization formats for the matrix and array2d objects so
        that either object can be deserialized into the other. This was done in a
        way that is backwards compatible with previous versions of dlib. That is,
        we can still load data serialized by previous dlib versions. However,
        older versions of dlib can't load the new serialization format.

Bug fixes:
   - Fixed a bug in save_dng() that happened sometimes when saving images with
     unsigned char pixels.
   - The test_ranking_function() and cross_validate_ranking_trainer() routines
     computed incorrect MAP scores when the learned function output a constant
     value for all samples.  This has been fixed.

Other:
   - Gave load_image_dataset() the ability to skip images that don't have any
     ground truth boxes.
   - Changed average_precision() to use interpolated precision. So now it uses
     the same metric as the one used by the Pascal VOC challenge.
   - Upgraded the dng image format so it can natively store floating point
     pixel types without any information loss.

New Features:
   - Added svd_fast(), a routine for computing a singular value decomposition of very 
     large matrices.
   - Added cca(), a routine for doing canonical correlation analysis on very large 
     and high-dimensional datasets.
   - Added tools for creating parallel for loops, see parallel_for().
   - Added some features to the image display widgets to let the user easily
     get information about where the user is clicking.  This is the new 
     get_next_double_click() routine.
   - Added an operator>> for matrix objects which allows you to read in ASCII
     matrices using the format used by operator<<.
   - Added serialization support for std::vector<bool>.
   - Added the following new minor objects and routines: average_precision(), 
     make_sparse_vector_inplace(), orthogonalize(), count_bits(), draw_surf_points(),
     hamming_distance(), cosine_distance, and negative_dot_product_distance.

Non-Backwards Compatible Changes:
   - Changed ranking evaluation functions to return the mean average precision
     in addition to just raw ranking accuracy. This changes their return types
     from double to matrix<double,1,2>.
   - Generalized segment_image() so it works on any pixel type or array of
     vectors.  I also changed its interface slightly.  In particular, I removed
     the min_diff parameter and replaced it with an explicit min_size parameter.
   - Changed how the SURF descriptor is computed slightly to improve its
     accuracy.  The interface to the user has not been changed, however, the
     number and position of detected SURF points might be different than in
     previous dlib versions.

Bug fixes:
   - Fixed an endianness bug in the PNG I/O functions which occurred when 16bit
     grayscale PNGs were used. 
   - Fixed a bug which could potentially occur when empty std::vector<char> or
     std::vector<unsigned char> were serialized.
   - There was a bug in the version of draw_line() that draws directly onto an
     array2d type image (not the one that draws onto a GUI canvas object). The
     bug triggered whenever a perfectly horizontal or vertical line that extended
     outside the image was drawn. This has been fixed.
   - Fixed a bug in the Windows implementation of the signaler object, which
     was found by Isaac Peterson. The bug caused the program to deadlock if
     signal() or broadcast() was called at exactly the same time a
     wait_or_timeout() function timed out.
   - Fixed a bug in the image_window and image_display GUI tools which caused
     them to not redraw overlay lines correctly in certain cases involving
     non-default zoom levels.
   - Switched randomly_color_image() to use the non-pointer based version of
     murmur_hash3() to avoid violation of the strict aliasing rule. In
     particular, the previous version didn't work correctly in gcc 4.7.2 when
     optimizations were enabled.
   - Visual Studio 2012's iostreams library has a bug which caused the
     iosockstream to crash on use.  This version of dlib has been changed to
     avoid triggering this bug.

Other:
   - Refactored the Platt scaling code a little. Now there is a function,
     learn_platt_scaling(), that allows you to directly call the Platt scaling
     code without supplying a trainer object.
   - Optimized the oca and structural SVM solvers.  They are now a little bit faster 
     than in previous dlib releases.

New Features:
   - Machine Learning
      - Added svm_rank_trainer, an optimized implementation of the SVM-Rank algorithm.
      - Added rank_unlabeled_training_samples(), an implementation of the SVM Active
        Learning algorithm.
      - Added svm_c_linear_dcd_trainer, a warm-startable SVM solver using the dual 
        coordinate descent algorithm used by liblinear.
      - Added the ability to force the last element of a weight vector to 1 to the
        following objects: svm_c_linear_trainer, svm_c_linear_dcd_trainer,
        svm_rank_trainer, and oca.
      - Added the ability to learn non-negative weight vectors to the
        structural_sequence_labeling_trainer object.
   - Networking
      - Added an iosockstream object.
      - Added a method to the server object that lets a user set the graceful close timeout
        time to something other than the default of 500ms.
   - Linear Algebra
      - Added the gaussian_randm() function.
      - Added the find_affine_transform() function.
      - Added the mat() function.  It combines the array_to_matrix(), vector_to_matrix(),
        pointer_to_column_vector(), and pointer_to_matrix() methods all into one convenient
        interface.   mat() also works for Armadillo and Eigen matrices.
      - Added STL style begin() and end() methods to matrix and matrix_exp.
      - Added an overload of sparse_matrix_vector_multiply() that multiplies a dense matrix
        with a sparse vector.
      - Made toMat() work with the matrix object in addition to array2d style images.
   - Graphical User Interface Tools
      - Added draw_solid_convex_polygon().
      - Added an overload of draw_image() that's useful for drawing images and doing
        interpolation at the same time.
      - Added the on_view_changed() callback to zoomable_region and scrollable_region widgets.
   - Added parse_trees_to_string() and parse_trees_to_string_tagged().
   - Added lambda function support to the timeout object.
   - Added the vectorstream object.
   - Added the parse_xml() routines.
   - Added a group name feature to the command line parser.  Now it is possible to make
     print_options() print related options in named groups. 
   - Added the following new hashing functions: murmur_hash3_128bit_3(),
     murmur_hash3_2(), murmur_hash3_3(), uniform_random_hash(), gaussian_random_hash() 
     as well as hash() overloads for uint32, uint64, and std::pair.

Non-Backwards Compatible Changes:
   - Made the svm_c_linear_trainer use the risk gap to decide when to stop.  This was done
     because it is how all the other OCA based SVM tools in dlib decide when to stop.  
     However, it might cause the outputs to be slightly different in this version of dlib.  
   - It is now illegal to call unlock() on a mutex when the mutex is not owned by the
     calling thread.  The most likely reason for doing this was to unlock early in an area
     locked by an auto_mutex.  Old code that does this can be fixed by calling auto_mutex's
     unlock() function instead.  
   - Removed the structural_assignment_trainer::learns_nonnegative_weights() routine 
     and moved its functionality into the feature extraction interface used by this object. 

Bug fixes:
   - Fixed a bug in find_max_factor_graph_nmplp() which caused it to not work properly on
     some compilers.
   - Fixed a bug pointed out by Joel Nelson in the version of md5() that took an istream.
     The bug caused the function to crash on strings longer than 56 characters.

Other:
   - dlib now has an excellent new logo thanks to Yasser Asmi. 
   - Added a new documentation page for the various linear algebra tools.
   - The following objects were turned into single implementation components:
     sockstreambuf, timeout, member_function_pointer, xml_parser, linker,
     bound_function_pointer, and timer.

New Features:
   - Machine Learning
      - Added the ability to learn non-negative weight vectors to the
        structural_assignment_trainer object.
      - Added two new graph clustering algorithms: Chinese Whispers and Newman's modularity
        clustering.
      - Added a number of new graph manipulation tools: sparse_matrix_vector_multiply(),
        is_ordered_by_index(), find_neighbor_ranges(), convert_unordered_to_ordered(),
        remove_duplicate_edges(), and the ordered_sample_pair object.
   - Networking
      - Added a set of tools for creating applications using the Bulk Synchronous Parallel
        computing model.  See the new bsp_ex.cpp example program for an introduction.
      - Added a routine that lets a user disable Nagle's algorithm on a TCP connection.
      - Added an asynchronous start routine to the server object.  This is the new
        start_async() method.
      - Added the network_address object.
      - Added connect_to() to the bridge interface.
   - Added find_max_parse_cky(), a method implementing the well known CKY algorithm for
     parsing probabilistic context free grammars.
   - Added the ability to label parts of objects with the mouse to the image_display
     widget.
   - Added the ability to put overlay circles and full_object_detections into the
     image_window widget.
   - Added a stddev() for matrix objects.
   - Added operator+() for running_stats and running_scalar_covariance.
   - Added an overload of murmur_hash3_128bit() that takes 4 integers instead of a block of
     memory.
   - Added rand::get_random_64bit_number().

Non-Backwards Compatible Changes:
   - Changed the image_dataset_metadata XML reading tools to use a map of strings to points
     to represent object parts. This change removes the old head point from a box since
     this information can now be represented in the parts map.
   - The syntax for passing order_by_distance and order_by_index to std::sort() is now
     slightly different since these functions are now templates.  However, this change
     allows them to work on any kind of sample_pair or ordered_sample_pair object.
   - The default distance value of a sample_pair is now initialized to 1 instead of
     infinity.

Bug fixes:
   - Added a patch, contributed by Martin Müllenhaupt, to fix a minor bug in the SQLite
     bindings.
   - Fixed a typo which would prevent code that called running_stats::max_n() from
     compiling.

Other:
   - Added a new documentation page for the various graph tools in dlib.
   - Added support for Visual Studio 2012.
   - Switched the sample_pair object to use double to store its distance value instead of
     float.
   - Added William Sobel's patch to the web server that improves its flexibility and
     security.
   - Changed the server object so you don't have to use the server::kernel_1a syntax to
     declare it anymore.  Now you just say server, server_iostream, or server_http
     depending on which one you want.
   - Changed the cmd_line_parser so you don't have to use the ::kernel_1a syntax anymore.
     Now it is declared like a normal single implementation object.
   - Set the default max number of connections a server will accept at a time to 1000
     rather than the previous default of infinity.

New Features:
  - Added more overloads of find_max_factor_graph_potts() to make applying it 
    over a Markov random field of image pixels really simple.
  - Added overloads of serialize()/deserialize() so that they can serialize
    Google protocol buffer objects.
  - Image Processing:
      - Added find_points_above_thresh()
      - Added max_filter() 
      - Added scan_image_movable_parts()
      - Added sum_filter_assign()
      - Added the full_object_detection object.
      - Added the ability to model objects with movable parts into the
        scan_image_pyramid object.  This update also includes all the needed tools
        to train movable part models using the structural_object_detection_trainer.
  - Machine Learning:
      - Added a per node loss option to the structural_svm_graph_labeling_problem's
        interface.
      - Added Emanuele Cesena's implementation of Sammon's nonlinear dimensionality
        reduction method.

Non-Backwards Compatible Changes:
  - To support movable part models, the serialization format of scan_image_pyramid
    objects was modified.  This breaks backwards compatibility with the previous
    format for scan_image_pyramid objects as well as object_detector instances
    that use the scan_image_pyramid.

Bug fixes:
  - Fixed a bug in auto_threshold_image() that caused it to give bad outputs 
    when used with very large images.

Other:
  - Updated find_max_factor_graph_potts() to correctly say you can use infinite
    weights for the factor_value_disagreement() values since the code actually
    supports this.
  - Made integer serialization about 3 times faster.

New Features:
   - Improvements to linear algebra tools:
      - Added the lowerbound() and upperbound() routines for thresholding dense 
        matrices.
      - Refined the tools for working with sparse vectors.  In particular, 
        the following functions were added: min(), max(), make_sparse_vector(),
        add(), and subtract().  A number of existing routines were also updated
        to work with both sparse and dense vectors so that templated code which 
        works on both vector types is simpler to write.
      - Added the += and -= operators to the set_subm(), set_rowm(), and set_colm()
        tools for operating on submatrices.
   - Optimization:
      - Added a new quadratic program solver, solve_qp4_using_smo().  This new
        solver is useful for solving quadratic programs corresponding to
        non-negative constrained primal quadratic programs.
      - Added an optional non-negativity constraint to the oca optimizer.
      - Added the min_cut object.  It provides a method to find the minimum weight 
        cut on a graph.  
      - Added tools for finding the maximum probability assignment in a Potts 
        style Markov random field.  See the find_max_factor_graph_potts() routine 
        for details.
   - Machine Learning:
      - Added structural SVM tools for learning the parameters of a Potts style
        Markov random field.  See the structural_graph_labeling_trainer and
        graph_labeler objects as well as their associated example program for 
        details.
      - Added the ability to learn only non-negative weights to the 
        svm_c_linear_trainer.
   - Improved Integration with OpenCV: 
      - Updated the cv_image object so it works with cv::Mat as well as IplImage.
      - Added the toMat() routine for converting from a dlib style image to an
        OpenCV cv::Mat image.

Non-Backwards Compatible Changes:
   - Removed the dlib::sparse_vector namespace.  Everything from this namespace 
     was moved into the normal dlib:: namespace so that code which works with 
     both sparse and dense vectors is more cohesive.

Bug fixes:
   - Fixed a bug in find_max_factor_graph_viterbi() which sometimes occurred when 
     the model order was larger than the number of variables.
   - Fixed a bug which caused a compiler error if you tried to call dot() on two 
     1x1 matrices which were statically dimensioned.

Other:
   - Improved existing documentation: added pictures of the gui widgets, 
     added documentation of the dlib::bridge protocol, and other minor 
     usability improvements.

New Features:
   - Image Processing:
      - Added the option to make the features generated by poly_image rotationally 
        invariant. 
      - Added a set of routines for warping, scaling, and resizing images.
        See the new "Scaling and Rotating" section of the image processing
        documentation for details.
      - Added the heatmap() routine for converting an image into a heatmap.
   - Machine Learning
      - Updated the sequence labeling trainer to allow the user to set different 
        loss values for different labels.
      - Added the rls object.  It is an implementation of the linear recursive 
        least squares algorithm.
   - Added the get_option() routines which slightly simplify option parsing 
     from the command line and config files.
   - Added the 128bit version of Murmur hash.
   - Added the kalman_filter and rls_filter objects.  These are tools for 
     performing Kalman filtering and recursive least squares filtering.
   - Added the circular_buffer object. 

Non-Backwards Compatible Changes:
   - The poly_image generates slightly different features in this new release.
     Therefore, classifiers trained using the previous version will need to be
     retrained if they are switched to the new version of poly_image.
   - Changed the xcorr() functions so they take the complex conjugate of the right
     hand arguments if they are complex numbers.  This way they do a proper
     cross-correlation and also mirror the behavior of MATLAB.  However, this
     breaks backwards compatibility with the previous behavior of xcorr().
   - Previously, dlib included two versions of dlib::array.  However, to
     simplify the user interface, dlib now includes only the contiguous
     memory implementation of dlib::array.  This change should only affect
     you if you wrote code which assumed dlib::array::set_max_size() only
     allocated a small amount of RAM.  The new behavior is similar to the
     std::vector::reserve() routine.  That is, dlib::array::set_max_size()
     will allocate the requested amount of memory immediately.

Bug fixes:
   - Fixed a bug which caused certain matrix expressions to not compile
     when the BLAS bindings were enabled.  In particular, expressions which
     involved a 1x1 matrix sometimes didn't compile.

Other:
   - Made the matrix routines min(), max(), sum() and mean() work with 
     complex numbers. 
   - Turned the array object into a single implementation object.  Now arrays 
     can be created using the normal array<type> obj; syntax.  Additionally, 
     all extensions were merged into the array object. 
   - Added an example program which better documents how to create training
     data for the object detection tools as well as how this data can be used.
     See the train_object_detector.cpp example for details.

New Features:
   - Added tools for timing blocks of code
   - Machine Learning
      - Added a set of tools for learning to solve the assignment problem.
        See the structural_assignment_trainer and its associated example
        program for an introduction.
      - Added random projection based locality sensitive hashing tools.
      - Added tools to simplify the creation of scan_image_pyramid objects.  
        See the object_detector_ex.cpp example program for details.
   - Image Processing
      - Added sum_filter() and spatially_filter_image_separable_down()
      - New feature extractors: poly_image, nearest_neighbor_feature_image, and
        fine_hog_image
 
Non-Backwards Compatible Changes:
   - Changed the serialization format for rand objects.
   - Changed the order of arguments for the sequence_labeler's constructor.
   - Object Detection Changes
      - Some parts of the object detection tools have been refactored.  In particular, 
        the interfaces of the scan_image_pyramid and structural_object_detection_trainer 
        have been changed slightly to improve usability.
      - Made the test_box_overlap a little more flexible.  This change breaks
        backwards compatibility with the previous version though.
      - The hashed_feature_image object has been made more general.  It now 
        uses a user supplied hashing function rather than its own hashing 
        implementation.
      - Removed constness from the operator() member functions of the 
        object_detector.
   - Fixed improper normalization in the gaussian() functions.  The
     normalization constant was being computed incorrectly.
   - Sequence labeling feature extractors must now define a sequence_type
     typedef instead of sample_type.  This change allows the user to use any 
     type of sequence, not just std::vector objects.

Bug fixes:
   - Changed the add_probability() method of joint_probability_table so
     it does a saturating add rather than a normal add.  This ensures the
     probability value stays exactly <= 1.  Previously, floating point
     rounding error could cause it to be slightly above 1 and would therefore
     cause some asserts to misfire during debugging mode.
   - The object_detector had code in it which limited the number of outputs
     to 100 rectangles.  This has been removed.
   - Fixed improper normalization in the gaussian() functions.  The
     normalization constant was being computed incorrectly.

Other:
   - dlib::rand can now generate Gaussian random numbers.
   - The structural_object_detection_trainer will now automatically setup
     appropriate non-max suppression parameters if the user doesn't supply them. 
   - The structural_object_detection_trainer has been optimized and now runs
     significantly faster than in previous dlib releases.
   - The tools folder containing htmlify, imglab, and mltool is now included
     in the dlib release archive files.  Previously, these tools were only 
     available directly from source control.

New Features:
   - Machine Learning
      - Added the histogram intersection kernel for sparse and dense vectors.
      - Added a set of tools to allow a user to easily learn to do sequence
        labeling using dlib's structural SVM implementation.  See the new
        sequence_labeler object and its associated example program for an 
        introduction.
   - Image processing:  
      - Added segment_image()
      - Added randomly_color_image()
      - Added the border_enumerator
   - Added the disjoint_subsets object, it is an implementation of the 
     union-find algorithm/disjoint-set data structure.
   - Added new matrix routines: conv(), conv_same(), conv_valid(), xcorr(),
     xcorr_same(), xcorr_valid(), and flip().
 
Non-Backwards Compatible Changes:
   - Changed find_max_factor_graph_viterbi() so you can use run-time
     defined order and num_states parameters.

Bug fixes:
   - The last dlib release added a max_iterations parameter to the
     svm_c_linear_trainer and svm_c_ekm_trainer objects.  However,
     there was a bug which made them only do at most 16 iterations,
     which is too few to solve many problems.  This has been fixed.
   - Fixed a bug in is_const_type.  It didn't work for reference types.
   - Fixed a bug in the SQLite binding routine statement::get_column_as_text().  
     It didn't work correctly if the column contained a NULL.
   - Fixed a bug in find_max_factor_graph_viterbi() which occurred when a
     zero order model had negative factor values.

Other:

New Features:
   - Two new routines for performing MAP inference in factor graphs:
      - For chain-structured graphs: find_max_factor_graph_viterbi()
      - For general graphs:          find_max_factor_graph_nmplp()
   - Image Processing
      - Added more tools for creating image pyramids.  See pyramid_down_5_4, 
        pyramid_down_4_3, and pyramid_down_3_2.
      - Added more image filtering and morphology functions.
      - Added a set of tools for creating sliding window classifiers:
         - Added the scan_image() routine.  It is a tool for sliding a set of 
           rectangles over an image space and finding the locations where the sum 
           of pixels in the rectangles exceeds a threshold.  Also added
           scan_image_pyramid, which is a tool for running scan_image() over an 
           image pyramid.
         - Added the structural_object_detection_trainer.  This is a tool which 
           formulates the sliding window classifier learning problem as an
           instance of structural SVM learning.
         - Added a variety of supporting tools and two object detection example 
           programs.
   - Added the following functions for computing statistics on vectors:
     mean_sign_agreement(), correlation(), covariance(), r_squared(),
     and mean_squared_error()
   - Added a C++ wrapper for SQLite (see the new database and statement objects)
 
Non-Backwards Compatible Changes:
   - Changed the interface to the ridge regression trainer objects so that they 
     report the entire set of leave-one-out prediction values rather than a 
     summary statistic like mean squared error.
   - Changed the serialization routine for bgr_pixels to store the pixels in BGR 
     order rather than RGB.
   - Changed the interface for the spatially_filter_image() routine to take the 
     filter as a matrix rather than C-array.  Also, now it won't force signed pixel 
     values to 0 if they go negative. 
   - Changed the test_regression_function() and cross_validate_regression_trainer()
     routines so they return both the MSE and R-squared values rather than just the
     MSE.
   - Changed suppress_non_maximum_edges() to use the L2 norm instead of L1 norm
     for measuring the strength of an edge since this produces a slightly better
     result.

Bug fixes:
   - The image_display didn't display overlay rectangles quite right.  If you zoomed
     in you could see that some of the pixels which are inside the rectangle were
     outside the overlay.  Specifically, the right column and bottom row was outside
     the overlay rectangle.  This has been fixed.  Now all pixels which are supposed
     to be part of a rectangle are drawn as being inside the overlay rectangle.
   - Fixed a bug pointed out by Martin Müllenhaupt which caused the windows socket
     code to not compile when used with the mingw-cross-env project.
   - Fixed a bug in the png_loader.  If you loaded an image with an alpha channel 
     into something without an alpha channel there were uninitialized values being 
     alpha blended into the image.
   - Fixed a bug in the cpp_tokenizer that only shows up on newer versions of gcc.  
     It wasn't tokenizing double quoted strings right.
   - Fixed a bug in spatially_filter_image() which showed up when using non-square 
     filters.  The bug would cause the edges of the output image to be incorrect.
   - Fixed a bug in the matrix class.  Expressions of the form mat *= mat(0) would
     evaluate incorrectly because the *= operator took the right hand side by reference
     and thus experienced an aliasing problem.  The other op= operators had similar
     problems and have also been fixed.
   - Fixed a bug pointed out by Justin Solomon which could cause the svr_trainer and
     svm_c_trainer to produce incorrect results in certain unusual cases.  

Other:
   - Added a more complete set of methods for converting between image space and 
     the downsampled hog grid used by hog_image.  Now you can convert from image 
     to hog in addition to hog to image.
   - Made the integral_image more general by making it templated on the type of 
     scalar used to store the sums.

New Features:
   - Added the check_sub_option() method to the command line parser check
     object.
   - Added match_endings to the dir_nav utils.
   - Added a set_current_dir() function.
   - Added the distance_to_rect_edge() routine.
   - Added support for user drawn rectangle overlays and selectable overlays 
     to the image_display widget.
 
Non-Backwards Compatible Changes:

Bug fixes:
   - Fixed a bug in the image_display widget.  If you switched it between
     images of a different size while any kind of zoom was in effect
     it could cause a segmentation fault.

Other:

New Features:
   - You can now add tasks to a thread_pool by value, using the new 
     add_task_by_value() method.
 
Non-Backwards Compatible Changes:

Bug fixes:
   - Fixed a bug which caused multiply defined symbol errors during linking
     if the PNG saving routine was #included.

Other:
   - Optimized the threaded and distributed structural svm solvers for the 
     case where there are many data samples and the separation oracle is 
     quick to evaluate.

New Features:
   - Added a function for saving to the PNG image format.
   - Added Austin Appleby's excellent MurmurHash3 hashing code and setup some
     additional convenience functions.  These functions are murmur_hash3() and
     various overloads of hash().
 
Non-Backwards Compatible Changes:
   - Made get_pixel_intensity() more efficient.  However, the value returned
     is slightly different than it used to be for RGB pixel types.

Bug fixes:
   - Setup proper error handling for libpng in the png_loader.  Now if the PNG
     file is corrupted in the middle it won't just print a message and abort
     the program.
   - Fixed a bug in assign_pixel_intensity() that happened when the target pixel 
     was an RGB pixel with an alpha channel.

Other:
   - Added a Frequently Asked Questions page
   - Changed the array2d object so you don't have to say array2d<type>::kernel_1a
     anymore to declare it.  Now you just say array2d<type>.

New Features:
   - Added tools for distributing the work involved in solving a structured
     SVM problem over many computers and CPUs. 
   - Added the bridge.  It allows a dlib::pipe to be used for networked
     communication.
 
Non-Backwards Compatible Changes:
   - Removed the DLIB_REVISION macro and replaced it with DLIB_MAJOR_VERSION and
     DLIB_MINOR_VERSION.

Bug fixes:

Other:
   - dlib's version control system has switched from Subversion to Mercurial.

New Features:
   - Added the max_sum_submatrix() function to the optimization tools.
   - Upgraded the pyramid_down function object so it can create color pyramids.  
     Also, added some functions which define the coordinate transforms between 
     different layers in an image pyramid.  
 
Non-Backwards Compatible Changes:
   - Changed the oca_problem interface to the oca optimizer.  Now the 
     optimization_status() function reports back the current risk and risk gap 
     in addition to the overall objective value and objective gap. 
   - Changed the stopping condition for the structured svm to the one suggested
     by the Joachims paper.  Now it stops when the risk gap is below a user
     supplied epsilon. 

Bug fixes:

Other:
   - Various usability improvements.
   - Improved the feature vector caching in the structural_svm_problem object.
   - Some objects were setup as multi-implementation objects but only had one 
     implementation.  I went through dlib and switched these to single implementation 
     objects.  So for example, to use the dlib crc32 module you used to declare an 
     object of type "crc32::kernel_1a" but now you can just say "crc32".  Note that 
     I did this change in a way that maintains backwards compatibility with previous 
     versions.  So crc32::kernel_1a is still allowed but that form is officially 
     deprecated.  The modified objects are as follows:
      - base64
      - byte_orderer
      - config_reader
      - crc32
      - pipe
      - rand

New Features:
   - Added a multiclass support vector machine.
   - Added a tool for solving the optimization problem associated with
     structural support vector machines.
   - Added new functions for dealing with sparse vectors: add_to(), 
     subtract_from(), max_index_plus_one(), fix_nonzero_indexing(), a
     more flexible dot(), and I renamed assign_dense_to_sparse() to assign() 
     and made it more flexible.
 
Non-Backwards Compatible Changes:
   - Renamed max_index_value_plus_one() (a function for working with graphs) to 
     max_index_plus_one() so that it uses the same name as the essentially 
     identical function for working with sparse vectors.
   - I simplified the cross_validate_multiclass_trainer(), cross_validate_trainer(), 
     test_binary_decision_function(), and test_multiclass_decision_function()
     routines.  They now always return double matrices regardless of any other 
     consideration.  This only breaks previous code if you had been assigning
     the result into a float or long double matrix.
   - Renamed assign_dense_to_sparse() to assign()

Bug fixes:
   - Fixed a bug in load_libsvm_formatted_data().  I had forgotten to clear the 
     contents of the labels output vector before adding the loaded label data.  
   - Fixed a bug in the kernel_matrix() function.  It didn't compile when used 
     with sparse samples which were of type std::vector<std::pair<> >.  
     Moreover, some of the trainers have a dependency on kernel_matrix() so this 
     fix makes those trainers also work with this kind of sparse sample.

Other:
   - Added a value_type typedef to matrix_exp so it's easier to write templates
     which operate on STL containers and matrix objects.

New Features:
   - Added an implementation of the Hungarian algorithm for solving the optimal 
     assignment problem (in the new max_cost_assignment() routine).
 
Non-Backwards Compatible Changes:

Bug fixes:
   - Fixed a problem which prevented the any_function unit test from compiling 
     in visual studio 2008.

Other:
   - Changed the oca optimizer so that it warm starts the QP subproblem
     rather than resolving it from scratch during each iteration. This 
     improves the speed and stability of the algorithm.

New Features:
   - Added the rr_trainer.  It is a version of krr_trainer which is optimized
     for use with the linear_kernel.
   - Added the approximate_distance_function() routine.  It is the core optimizer
     behind the reduced2() trainer adapter.  
   - Added an any_function which supports the same functionality as std::function 
     from the upcoming C++0x standard.  I added this so dlib can be modified to 
     easily support lambda functions while still being compilable with compilers 
     which don't support the new std::function.
   - Added overloads of all the GUI event handlers so you can use general functions 
     as callbacks (via any_function).  This way, if you have a C++0x compiler, you 
     can use lambda functions with the event handlers.
   - Added the split() function for splitting up strings.
 
Non-Backwards Compatible Changes:
   - Improved the distance_function object by turning it into a properly 
     encapsulated class rather than just a simple struct.  I also added 
     overloaded +, -, *, and / operators for this object so you can do the 
     kind of arithmetic you would expect on an object which represents a 
     point in a vector space.  This breaks backwards compatibility with 
     the previous interface though as the member variables are now private.

Bug fixes:
   - Fixed a compile-time error in the kernel_matrix().
   - Fixed a bug in an assert in the spatially_filter_image() function.
   - Applied a patch from Nils Labugt which fixes a runtime bug in the gui_core 
     component.  The bug caused a crash when using X11 and Ubuntu 10.10 in
     certain cases.
   - Updated code so that it compiles with the clang compiler.

Other:
   - Updated the image_display widget so you can zoom in and out using the 
     mouse wheel.

New Features:
   - General Stuff 
      - Added the promote template
      - Added the basic_type template
      - Added the assign_image_scaled() function
      - Added the unordered_pair object.
      - Added the symmetric_matrix_cache() function
      - Added two new quadratic program solvers.  The solve_qp2_using_smo
        and solve_qp3_using_smo objects.

   - Machine Learning Stuff
      - Added the svm_c_trainer object.  It implements C-SVM classification and
        allows the user to set different C values for each class.
      - Added the svm_one_class_trainer object. 
      - Added the svr_trainer object.  It implements epsilon-insensitive
        support vector regression.
      - Added two new any objects.  The any_decision_function for containing 
        decision function style objects and the any_trainer for trainers.
      - Added cross_validate_regression_trainer()
      - Added test_regression_function()
      - Added the probabilistic() function.  It is a trainer adapter that
        simply calls train_probabilistic_decision_function().
      - Added tools for multiclass classification 
         - Added one_vs_one_trainer
         - Added one_vs_all_trainer
         - Added cross_validate_multiclass_trainer()
         - Added test_multiclass_decision_function()
 
Non-Backwards Compatible Changes:
  - invalid_svm_nu_error has been renamed to invalid_nu_error.
  - Changed the pixel_traits so that signed grayscale pixel types are allowed.  
    This involved adding a few new fields to pixel_traits.  I also changed the 
    get_pixel_intensity() function so that its return value is of the same type 
    as the basic pixel type rather than always being unsigned long.
  - Removed the kernel_type typedef from the normalized function since this
    meta-object should be capable of working with non-kernel decision functions.
  - train_probabilistic_decision_function() no longer accepts column vectors of 
    samples and labels.  Now it only accepts std::vectors of samples and labels.  

Bug fixes:
  - Fixed a bug in the deserialization code for the sparse kernels.  The bug
    prevented code which used the deserialize() routine from compiling.

Other:
  - Changed the image display GUI widgets to use the assign_image_scaled() 
    function internally.  Now they will accept just about any image and 
    do the right thing.
  - Modified the type_safe_union so that you can call apply_to_contents() on const
    type_safe_unions.
  - Added serialization support for std::pair objects.
  - Made the train_probabilistic_decision_function() more general by making it work 
    with any kind of trainer object rather than only ones which produce 
    dlib::decision_function objects.  I also made it work with trainers that only 
    take std::vectors.
  - Added overloads to the config_reader's methods to allow it to load directly 
    from a file name given as a string in addition to taking istream objects.

New Features:
  - Added the ability to add/subtract scalar values to/from all the elements 
    of a matrix using the - and + operators.
  - Added a trust region optimizer. 
  - Added Levenberg-Marquardt and LM/quasi-newton hybrid methods for solving
    non-linear least squares problems.
  - Added an any container object.

Non-Backwards Compatible Changes:

Bug fixes:
  - Fixed a compiler warning and also a runtime bug in sort_basis_vectors().  
    The bug triggered when all the basis vectors were included in the final
    answer.

Other:
  - Added a bunch of overloads to catch operations on diagonal matrices
    and use more efficient code paths for them.  For example, inv(diagm(d))
    turns into diagm(reciprocal(d)).  Multiplication by a diagonal matrix
    is now also handled efficiently.

New Features:
  - Added a class for reading JPEG image files.
  - Added scale_rows(), flipud() and fliplr() matrix functions.
  - Added console_progress_indicator.  It is a tool for measuring how long a 
    task will take.
  - Added sort_basis_vectors().  It is a function for performing supervised 
    basis selection.

Non-Backwards Compatible Changes:
   - Renamed the linearly_independent_subset_finder's dictionary_size() member 
     function to size().  This way, linearly_independent_subset_finder objects 
     can be used in many templated functions which expect objects which look 
     like arrays. 

Bug fixes:
  - Changed the assert macros so that they don't use __PRETTY_FUNCTION__
    with gcc 4.4.5 since, on Ubuntu at least, this version of gcc segfaults
    when __PRETTY_FUNCTION__ is used within certain template constructs.
  - Improved the alias detection capability of kernel_matrix() expressions.  
    Now statements of the form: sample = kernel_matrix(kern, *, sample) can 
    be used since the aliasing of sample will be handled.

Other:
  - Generally tried to make things more usable.
  - Optimized matrix expressions such as mat*diagm(vect)
  - Made the code in chol() more robust to indefinite matrices.

New Features:
   - Added the running_scalar_covariance object.
   - All the matrix decomposition routines now use LAPACK when DLIB_USE_LAPACK
     is #defined.   

Non-Backwards Compatible Changes:
   - Removed the dlib::EOTHER constant since it conflicts with visual
     studio 2010.
   - Changed the svd functions so you can't supply output matrices which use 
     both column and row major layouts.  Now all the output matrices need to 
     use the same memory layout.
   - Removed the qr_decomposition::get_householder() function.

Bug fixes:
   - Minor fixes so that dlib compiles in Visual Studio 2010

Other:
   - Added an overloaded matrix_assign() that handles symmetric kernel_matrix()
     expressions more efficiently by only evaluating the upper triangular part
     of the matrix.

New Features:

Non-Backwards Compatible Changes:

Bug fixes:
   - Fixed a compile-time bug in the matrix related to multiplication by 
     subm() expressions when the BLAS bindings were enabled.
   - Fixed a bug in train_probabilistic_decision_function() which could 
     cause it to go into an infinite loop when working with very large 
     datasets.

Other:

New Features:
  - Added a reference_wrapper implementation and modified the thread_function 
    slightly so it works with it.
  - Added an implementation of kernel ridge regression.
  - Added a simple newton search strategy for optimizing functions.

Non-Backwards Compatible Changes:
  - If you have created your own matrix expressions then its possible this
    new release will cause them to not compile.  

Bug fixes:
  - Fixed a bug in scale_columns.  It said it didn't have any destructive aliasing 
    when in fact it destructively aliased its second argument.
  - Fixed a bug in the random number generator where setting the seed back to ""
    didn't result in the object going back to its initial state.

Other:
  - Reorganized the matrix expression code.  It's now much simpler and the
    library includes a new example program which details the steps needed to
    create new matrix expressions.
  - Changed the train_probabilistic_decision_function() routine so that it uses
    a more numerically stable method to perform its maximum likelihood optimization.
  - Added missing get/set epsilon functions to the RVM training objects.
    I also changed the default epsilon from 0.0005 to 0.001.

New Features:
   - Added the simplify_linear_decision_function() routines.
   - Added the find_approximate_k_nearest_neighbors() function.
   - Added the fill_lisf() function.

Non-Backwards Compatible Changes:
   - Made the sample_pair have a default distance of infinity instead of
     the max floating point value.  I also reworked the graph creation functions 
     to make them a little more versatile.  Now you can use infinite distances to 
     indicate that certain nodes are not connected at all.
   - Changed the linear_manifold_regularizer to normalize the regularization
     parameter by the sum of edge weights instead of the sum of edges.

Bug fixes:
   - Fixed a bug in the timer_kernel_2 object.  In certain rare cases it would 
     stop calling the action function and essentially shut down without being 
     told to do so.

Other:
   - Made the reduced() and reduced2() functions more efficient.
   - Many small usability improvements here and there.

New Features:
   - Added the svm_c_ekm_trainer.  It is a kernelized version of the fast
     linear trainer svm_c_linear_trainer.
   - Added the linear_manifold_regularizer and some supporting tools.
   - Added the sum_rows(), sum_cols(), join_rows(), join_cols(), reshape(), 
     and pointer_to_matrix() functions.
   - Added the randomly_subsample() function.

Non-Backwards Compatible Changes:

Bug fixes:
   - Fixed some minor compile time bugs on certain older compilers.

Other:
   - Updated the += and -= matrix operators to be a little more flexible.  Now
     if you try to apply them to a matrix of the wrong size it will automatically
     resize the target matrix and just do a normal assignment.
   - Removed the requirement that you load dng files into an image of the exact
     pixel type that created the file.  Now you can use any pixel type.  I also
     changed the code so that grayscale pixels with more than 16 bits get saved as
     16 bit grayscale images instead of 8 bit images.

New Features:
   - Added the solve_qp_using_smo() function to solve certain quadratic
     programs.
   - Added the oca object.  It is an implementation of the Optimized Cutting 
     Plane Algorithm.
   - Added a linear SVM trainer that uses oca.
   - Added an implementation of the Histogram of Oriented Gradients algorithm
   - Added a simple tool for making image pyramids
   - Added the running_covariance object 
   - Added a simple linear (i.e. non-kernelized) kmeans implementation
   - Added support for serializing dlib::int64
   - Added some functions to load and save LIBSVM formatted data files.

Non-Backwards Compatible Changes:
   - Changed the definition of dlib's sparse vector format to require
     unsigned integral keys.  Having this requirement is nice because it 
     creates a simple correspondence between dense vector index values and 
     sparse vector keys.  The previous sparse vector definition was  
     excessively generic.
   - Renamed sparse_vector::dot_product() to sparse_vector::dot() so that 
     both dense and sparse vectors have a global function with the same 
     name (i.e. dot()).

Bug fixes:
   - Fixed a bug discovered by Mitchell Peabody.  In some instances trying to
     deserialize a std::vector would fail to compile.  

Other:
   - Increased the number of template arguments of the type_safe_union from 10 
     to 20.  Additionally, I made the get_id() function public and renamed it 
     to get_type_id().  I also added a comment explaining the serialization 
     format of type_safe_union objects.
   - Moved the optimization algorithms into their own page in the documentation.
   - Added a Suggested Books page to the documentation

New Features:
   - Added the ability to compute transformation matrices that map between
     the representations used by different empirical_kernel_maps.  Also added
     the ability to compute projection error.
   - Added the random_subset_selector object.
   - Added the compute_mean_squared_distance() function.

Non-Backwards Compatible Changes:
   - Modified the logger's hook implementation so that it uses a special stream  
     buffer instead of an std::ostringstream.  This way logging doesn't cause 
     memory allocations.  This breaks backwards compatibility with the previous
     hook function API but only slightly.  The new hook functions must take a
     const char* instead of std::string.
   - Added the const_ret_type typedef to the matrix_exp.  It is now required that 
     all matrix expressions define this type.  This enables the expressions to 
     return elements by constant reference when appropriate rather than always 
     returning by value.  

Bug fixes:
   - Fixed a bug in the matrix BLAS bindings that caused BLAS to return an invalid
     argument error.  The error occurred when general matrix multiply expressions
     were transposed and didn't result in a square matrix.  E.g. mat = trans(a*b)
     where mat isn't square.
   - Fixed potential compile time bugs in the comparison operators for futures.
   - Added a missing check for division by zero in the SURF feature extractor.
   - Modified the find_min_single_variable() function so that it is more 
     robust when working with functions that are made up of a bunch of
     constant value strips.  Previously, these kinds of functions could
     cause the optimization to fail.

Other:
   - Changed the regression test suite so that when it sets the logging level 
     it now sets it for all loggers.  Not just ones that start with "test."

New Features:
   - Added some MATLAB style thresholding relational operators to the matrix.
   - Added the kernel_matrix() functions.
   - Added the empirical_kernel_map object.
   - Added the discriminant_pca object.
   - Added the read_write_mutex object.

Non-Backwards Compatible Changes:
   - Renamed the support_vectors member of the decision_function and
     distance_function classes to basis_vectors.  This name more appropriately 
     reflects how these two classes are used within the library.
   - Changed the matrix_exp interface slightly.  This could only impact users
     who created their own custom matrix expressions.  If you don't get a 
     compiler error then you don't have to worry about it.

Bug fixes:
   - Fixed a minor error in the LBFGS code.
   - Added a missing check for division by zero to the kcentroid, krls,  
     and linearly_independent_subset_finder classes.  If someone added
     the zero vector to them as the first training example a division by zero 
     could result.
   - There were a few cases where the code wouldn't compile when using
     matrices of complex numbers.  There was also a runtime bug that triggered 
     when a rank 1 update was performed where one of the vectors was conjugated
     and two or more transposes were used in certain positions.  This bug
     caused the wrong output to be computed if the BLAS bindings were used.  
     Both of these bugs have been fixed.
   - Fixed a bug in the http server that affected cookies with certain kinds of 
     data.  The result was invalid data being sent back to the web browser.

Other:
   - Generally improved the BLAS bindings for the matrix object.

New Features:
   - Added the pointer_to_column_vector function.
   - Added the BOBYQA algorithm for derivative-free optimization.
   - Added some functions to make it easy to do a line search on a function
     of a single variable when derivatives are not available.

Non-Backwards Compatible Changes:

Bug fixes:
   - Fixed a bug in the cpp pretty printer.  It wasn't parsing 
     exponentiated numbers like 1e100 correctly.

Other:
   - Added a model selection example program using grid search
     and the new BOBYQA algorithm.

New Features:
   - Added an implementation of the L-BFGS algorithm for unconstrained non-linear
     optimization. 

Non-Backwards Compatible Changes:
   - Refactored the optimization code.  It is now much more flexible but
     this resulted in changes to the API.  See the optimization example program 
     for a discussion of the new API.

Bug fixes:
   - Fixed a bug in the get_filesystem_roots() roots function that
     prevented it from compiling.

Other:

New Features:
   - Added the ability to use a kernel cache to the batch_trainer object.  
   - svm_pegasos can now be configured to use two different lambda arguments
     for use with unbalanced data.
   - Added the reciprocal_max() and dot() matrix functions.
   - Added the bgr_pixel and cv_image objects so that OpenCV images can
     be easily used with dlib routines.

Non-Backwards Compatible Changes:
   - I changed the batch trainers so that they always call clear() on the 
     trainer being used before training begins.  
   - Modified the svm_pegasos class so that the user can set independent lambda
     parameters for each class.  This breaks backwards compatibility with 
     the previous interface slightly and changes the serialization format
     of this class.
   - Split the vector_normalizer into a normal normalizer and a pca normalizer
     version. 
   - The zoomable_region widget now uses exponential rather than linear 
     zoom scaling since this is much more pleasing to use.  There is now 
     a new requirement on the zoom increment that it must be between 0 
     and 1.

Bug fixes:
   - Fixed a bug in the cross_validate_trainer_threaded() function.  It could 
     deadlock if more than about 10 folds were requested.
   - Fixed the serialization functions for the normalized_function object.  
     They will now work with custom normalizer function objects.
   - Fixed a minor bug in the zoomable_region::set_min_zoom_scale() function.
     It didn't always end up zooming in a smooth sensible manner after this
     function was called.

Other:
   - Made the thread_function object more general.   It can now handle
     arbitrary functions of up to four arguments.

New Features:
   - Added the reshape_to_column_vector() function.
   - Added a hook to the logger object that lets you set a different kind of
     output logging destination (in addition to the std::ostream supported 
     already).
   - Upgraded the scoped_ptr so that it can handle array pointers as well
     as customer deleter functions.
   - Added overloads of the kernel_derivative object for all the kernels 
     in dlib.

Non-Backwards Compatible Changes:
   - Reworked the config_reader interface a little to make it easier to use.  
     In particular, I removed the enumerator over blocks in favor of a simple 
     get_blocks() function that just returns a std::vector of all the blocks.  
     I also removed the requires clauses on the block and key accessor functions 
     and instead made a request for a non-existent key/block result in a non-fatal 
     exception.  This way users can let the config reader perform a more natural 
     role in config file validation (by catching this exception and acting 
     accordingly).
   - It is now illegal to multiply matrices of size zero together. 

Bug fixes:
   - Fixed the gaussian() function used by the SURF code.  It wasn't computing 
     a properly weighted Gaussian function.
   - Fixed a few things in various parts of the code to avoid compiler errors 
     in certain use-cases.
   - Added a missing rethrow statement.  The xml parser would eat exceptions 
     thrown by event handlers rather than letting them propagate out as 
     documented in the specification.

Other:

New Features:
    - Added an implementation of the SURF algorithm which includes the
      following new objects and functions:  integral_image, hessian_pyramid, 
      interest_point, surf_point, compute_dominant_angle(), 
      compute_surf_descriptor(), haar_x(), haar_y(), get_interest_points(), 
      and get_surf_points().
    - Added the zeros_matrix() and ones_matrix() functions.
    - Added serialization support to the type_safe_union object.
    - Added the grow_rect() and shrink_rect() functions.
    - Added the get_files_in_directory_tree() function.
    - Added the null_trainer_type object.
    - Added the roc_trainer_type object.

Non-Backwards Compatible Changes:
    - Removed some extraneous get_kernel() functions from some of the 
      trainer adapter classes since they really aren't needed.  

Bug fixes:
    - Changed the socket read/write code so that it can handle a large 
      number ( > 2 billion) of bytes in I/O calls.
    - Added a missing type cast to the reciprocal() function to fix a compile
      time error you get when you use it with complex<float> type matrices.
    - Fixed a bug in the assign_border_pixels() and zero_border_pixels() functions.  
      Their contracts said there was no upper limit on the size of the border that 
      could be assigned/zeroed but the implementations failed to handle the case 
      where the border was bigger than the image.

Other:
    - Generally cleaned up the code and documentation here and there.
    - Added in Steven Van Ingelgem's patches to improve the usability of the
      HTTP server object.
    - Updated the load_bmp() function so that it is capable of reading BMP
      files that have been compressed with the RLE compression sometimes
      used for 8bit BMP files.
    - Merged in Miguel Grinberg's patch to add a non-blocking read() function to the
      connection object.

New Features:
   - Added a set of kernels that can operate on sparse vectors.
   - Added the image_window and image_display objects.
   - Added the rotate_point() function and the point_rotator object.

Non-Backwards Compatible Changes:
   - Added Steven Van Ingelgem's patch to add the body of data posted 
     back to the server into the incoming data object given to the 
     server_http::on_request() handler.  This removes the content_length 
     field and replaces it with a string that contains the body of content
     data.

Bug fixes:
   - Fixed a compile time bug in the offset_kernel.

Other:
   - Added optimized overloads of the kcentroid object for various 
     linear kernels.
   - Changed all the tests in the dlib test suite to use a new DLIB_TEST 
     macro instead of DLIB_CASSERT since the tests really aren't 
     technically assertions

New Features:
   - Added the strings_equal_ignore_case() functions

Non-Backwards Compatible Changes:
   - Changed the on_request() function in the http server
   - Changed the serialization format of the kcentroid and svm_pegasos
     objects
   - By default, the kcentroid now keeps the most linearly independent 
     dictionary vectors rather than the newest

Bug fixes:

Other:
   - Split the algorithms documentation page into three pages, algorithms,
     machine learning, and bayes nets.
   - Merged in Steven Van Ingelgem's patch to cleanup the HTTP server and
     add new functionality.  This breaks backwards compatibility with the
     previous on_request() interface but it is easy to update old code and
     it is now much cleaner and easier to use.
   - Changed the kcentroid so that you can tell it to keep the most linearly 
     independent vectors rather than the newest vectors.  I then changed the
     svm_pegasos object so that it has a max number of support vector setting
     so that the user can supply an upper limit on the number of support 
     vectors to use.  

New Features:
   - Matrix related
      - Added the find_min_and_max(), index_of_min(), index_of_max(), trace(),
        randm(), linspace(), logspace(), and cartesian_product() functions.
   - Machine learning related
      - Added the offset_kernel
      - Added some functions to the kcentroid to allow the user to compute 
        the inner_product of kcentroids as well as a few other useful things.
      - Added a kernelized version of the Pegasos SVM training algorithm.

Non-Backwards Compatible Changes:
   - Changed the range() function so that it returns row vectors
     instead of column vectors.  

Bug fixes:
   - Changed threading code to avoid a potential race condition during 
     program termination.
   - Fixed a few incorrect DLIB_ASSERT statements
   - Fixed a bug in the way Content-type was handled in HTTP posts.
   - Fixed a bug in subm() that showed up when statically dimensioned row 
     vectors were used to select a sub matrix.

Other:
   - Added some functions to the rectangle to make it easy
     to get the corner points.
   - The cross validation functions no longer allow invalid_svm_nu_error 
     exceptions to escape.  Instead, they are assigned low CV scores.
   - Made std_vector_c able to copy directly from std::vector objects.
   - Added a get_socket_descriptor() function to the connection class.

New Features:
   - Matrix related
      - Added QR, LU, Cholesky, and eigenvalue decomposition class objects
      - Added overloads for rowm() and colm() that allow you to pick out
        less than an entire vector
      - Added the lowerm() and upperm() functions
      - Added the const_temp_matrix class

Non-Backwards Compatible Changes:
   - Renamed the cholesky_decomposition() function to chol()

Bug fixes:
   - Fixed some errors in the requirements for calling the new rowm() and 
     colm() functions.
   - Fixed dlib::abs() so that it returns the right type when used
     with complex matrices.
   - Fixed a race condition in the logger object.  It was missing a needed call
     to unregister_thread_end_handler().  What could happen in some scenarios is,
     during program termination, a global part of the logger object could be destructed
     when it still had outstanding thread end handlers registered to it.  

Other:
   - Added an example program that shows how to use the optimization 
     functions.
   - Gave the matrix object the ability to factor expressions containing
     trans() function calls into more efficient forms.
   - Generally cleaned up the matrix code 

New Features:
   - Added the multi-line text_box GUI widget.
   - Added the type_safe_union object

Non-Backwards Compatible Changes:
   - Renamed the array::expand() function to array::resize() since it does 
     basically the same thing as std::vector::resize() and more than one
     user has told me they found the name "expand" to be confusing.

Bug fixes:

Other:
   - Added an example showing how to use the type_safe_union and pipe
     together.
   - Added a page to the documentation that discusses the dlib coding
     standards and how to contribute to the project.

New Features:
   - Added the bound_function_pointer object.
   - Added support for futures to the thread_pool object.
   - Added a set of objects that makes it possible to create simulations
     of quantum computing algorithms.
   - Added copy and paste support to the text_field.
   - matrix object stuff
      - Added the range() function as well as overloads of all the various 
        sub-matrix selection functions so that you can pick out slices of 
        matrices like in Matlab.
      - Added a new template argument to the matrix object that allows the
        user to select the memory layout.  Also added a row_major_layout 
        and column_major_layout.
      - The matrix object can now be initialized using a comma separated
        list of values.

Non-Backwards Compatible Changes:
   - Changed the fatal_error exception class so that it aborts your program
     and prints a message if you try to construct it more than once since 
     doing so indicates that you ignored the first fatal error.
   - The way matrix expressions work in the library has been changed
     since the last release.  So if you created custom matrix expressions 
     then they will need to be updated to use the new matrix expression stuff.

Bug fixes:
   - Fixed a minor bug in how the zoomable_region widget drew itself after 
     a resize in some cases.
   - Fixed a problem with draw_line where it didn't always redraw the line 
     properly.

Other:
   - A lot of the matrix code has been refactored and optimized.  The matrix
     object will now introduce temporary objects when doing so results in
     better performance.  I also added a simple system for binding 
     arbitrary matrix expressions to optimized BLAS routines. 
   - Cleaned up the vector and point classes.  Now there is only one class, 
     the vector class, and it is capable of representing everything the old 
     vector and point class could.  I also added code to make sure the 
     vector class does the appropriate type promotions when vector objects
     with different types are used together.
   - Made the vector class inherit from matrix

New Features:
   - Added user settable styles to most of the gui widgets
   - Added the diagm(), svd2() and svd3() matrix functions
   - Added the thread_pool object

Non-Backwards Compatible Changes:
   - Removed the arrow_button widget and moved its functionality into the 
     button widget. 
   - Renamed the dragable class to draggable 
   - Removed the confusing and unnecessary hidden bool argument to the 
     gui widget style drawing functions.
   - Changed some of the events that are about the mouse leaving a widget so 
     that they still trigger even if the widget has been disabled or hidden.  

Bug fixes:
   - Added some missing mutex locks to the scroll_bar widget
   - Fixed a bug in the fill_gradient_rounded() function. It didn't always
     draw the entire rectangle.
   - Fixed a compile time bug in the pinv() function.  It didn't compile 
     when used on statically sized matrices when they weren't square.

Other:
   - The member_function_pointer object now never calls new or delete. 
     So it is safe to use in a real time environment.

New Features:
   - Added the sort_columns() and rsort_columns() functions
   - Added the vector_normalizer object
   - Added the normalized_function object.
   - Added a tensor_product() function for the matrix object.

Non-Backwards Compatible Changes:

Bug fixes:
   - Made it so that the gui event handler thread isn't created at all 
     unless some part of an application calls some of the gui_core code. 
     In the previous release the event handler thread was executed
     briefly during program termination and could cause problems if no
     windowing environment was available.
   - Fixed an #include statement in the matrix utilities so that it works 
     even if you don't specify an include path argument to your compiler.

Other:

New Features:
   - Added a thread safe shared pointer object
   - Added the popup_menu_region widget.

Non-Backwards Compatible Changes:
   - The on_wheel_up() and on_wheel_down() gui events now take an unsigned long
     argument.  
   - Removed the register_program_ending_handler() function from the threading
     API and also changed the dlib thread pool so that it no longer causes
     a terminating program to wait for any outstanding threads to finish
     before allowing the application to end.
   - Changed the serialization format of the linearly_independent_subset_finder
     class.
   - Changed all the font pointers in the gui API's interfaces
     to shared_ptr_thread_safe objects.

Bug fixes:
   - Made the kkmeans class actually use the min_change parameter.
   - Fixed a bug in the linearly_independent_subset_finder object.  Also
     added a way to set a minimum tolerance.  
   - Fixed a bug in the scrollable_region widget that caused it to scroll in an
     unpleasant way when the horizontal and vertical scroll increments weren't 
     set to the same value.
   - Made one of the arguments to font::draw_string() not be a reference because
     some versions of gcc don't end up doing the right thing when -O3 is 
     supplied.
   - Fixed a bug in the covariance() function that prevented it from compiling
     sometimes.

Other:
   - Changed the gui core code around so that it should be safe to make window 
     objects at the global scope.
   - Added more control over how the scrollable_region scrolls its region.
     You can now adjust how much it scrolls when the mouse wheel is scrolled 
     as well as enabling scrolling via a mouse drag.
   - Modified the library so that it compiles with the Intel compiler.
   - Added some example programs that use the relevance vector machine

New Features:

Non-Backwards Compatible Changes:

Bug fixes:
   - Fixed a bug in the tooltip widget
   - Fixed the cmake option to toggle the ENABLE_ASSERTS macro
   - Fixed some bugs in the rvm
   - Fixed the serialization code for the kkmeans object so that it actually 
     works
   - Fixed a bug that can trigger when the thread_specific_data object is 
     destructed
   - Fixed a bug in the directory navigation gui.  If you tried to go
     into a drive on windows that wasn't mounted you got an error.
     This is now fixed.

Other:
   - Made the dir_nav stuff work with std::vector and dlib::std_vector_c 
     as well as dlib::queue objects.
   - Generally cleaned up a bunch of things

New Features:
   - Added relevance vector machine regression and classification support.
   - Added the cross_validate_trainer_threaded() function
   - Added the length and length_squared matrix functions.

Non-Backwards Compatible Changes:

Bug fixes:
   - Changed gui code a little so that windows don't popup in funny 
     places when used with the cygwin X windows system.

Other:
   - Made it easier to use the scoped_ptr with the TCP connection object
   - Optimized the matrix object a little

New Features:
   - Machine Learning
      - Added the ability to compare kcentroid objects to each other
      - Added the rank_features() function
      - Added the distance_function object
      - Added the reduced_decision_function_trainer object and 
        reduced() function
      - Added the reduced_decision_function_trainer2 object and 
        reduced2() function
      - Added a radial basis function network trainer
      - Added the linearly_independent_subset_finder object
      - Added the sigmoid_kernel
   - Matrix Utilities
      - Added the inv_upper_triangular() and inv_upper_triangular()
        functions.

Non-Backwards Compatible Changes:
   - Refactored a bunch of the kernel learning code into a much cleaner form.
     But this does change the interface to all the training functions.  

Bug fixes:
   - Fixed a bug in the min and max calculation in the running_stats object
   - Removed a bug in the sum() and variance() functions that
     caused them to seg fault when they were used on certain 
     matrix of matrix objects.
   - Added a missing check for division by zero to the conjugate gradient
     optimization functions.
   - Added some missing member variables to the .swap and serialization 
     functions for the kcentroid object.  So now they should work right.

Other:
   - Added an option to the cmake file to toggle the DLIB_ASSERT macro
   - Added an option to the cmake file to toggle the dlib stack trace macros
   - Made the library compile in Cygwin

New Features:
   - Merged in Keita Mochizuki's unicode patches to the GUI components.  So
     the dlib GUI now has better unicode support.
   - Added the remove_row and remove_col matrix functions.  Also made all
     three of the above functions capable of taking arguments at run time 
     as well as compile time.
   - Added the ability to cap the number of dictionary vectors used by the krls
     and kcentroid object at a user specified number.   
   - Added the pick_initial_centers() function
   - Added the running_stats object

Non-Backwards Compatible Changes:
   - Changed the interface to the krls and kcentroid objects somewhat.
   - All of the style objects in the GUI part of the library now use 
     dlib::ustring instead of std::string.  This only matters to you if 
     you have made your own style objects.
   - Changed the serialization format of the krls, kcentroid, and
     directed_graph_drawer objects.  Note that is also means that the
     files saved by previous versions of the bayes_net_gui_ex program
     won't load with the current version.

Bug fixes:
   - Fixed an aliasing bug in the set_subm(), set_rowm(), and set_colm() 
     functions.  It was possible that you could get incorrect results
     if you used these functions to copy one part of a matrix to another
     part of the same matrix if the two areas overlapped.
   - Fixed a minor numerical error in the krls code so now it gets slightly 
     better results.

Other:
   - Made the types generated by the matrix operations a lot shorter.  This
     avoids some compiler warnings from visual studio and even some potential
     internal compiler errors in some instances.

New Features:
   - Added some macros that allow dlib to create a stack trace
   - Added the wrap_function and is_function templates.
   - Added two new events to the text_field object.  One for detecting when the 
     user hits enter and another for detecting when input focus is lost.
   - Machine Learning
      - Added a kernel based centroid estimator/novelty detector
      - Added a kernel based k-means clustering algorithm
   - Numerical
      - Added an identity_matrix() function that can take a runtime defined size.
      - Added a bunch of unconstrained optimization stuff to the library. 
        It now has a conjugate gradient optimization algorithm as well as
        a quasi-newton algorithm.
      - Added the set_subm, set_rowm, and set_colm functions.

Non-Backwards Compatible Changes:
   - In the krls object:  Added a requires clause to the set_tolerance() member 
     function and renamed clear() to clear_dictionary().

Bug fixes:
   - Fixed a bug in the requires clause of the subm() function.  It was 
     more restrictive than it should have been.

Other:
   - Added example programs for the krls object as well as the new
     kcentroid and kkmeans objects.

New Features:
   - Added an implementation of the kernel recursive least squares algorithm

Non-Backwards Compatible Changes:
   - Broke backwards compatibility in the directed_graph_drawer's serialization
     format when I fixed the bug below.  

Bug fixes:
   - Fixed two bugs in the directed_graph_drawer widget.  First, it sometimes 
     threw a dlib::fatal_error due to a race condition.  Second, the color of 
     the nodes wasn't being serialized when save_graph() was called.  
   - Made vector_to_matrix() work for std::vector objects that have non-default 
     allocators.

Other:
   - Added some stuff to make people get a really obvious error message
     when they set up the include path incorrectly.

New Features:
   - Added the vector_to_matrix() function.
   - Added a cholesky_decomposition() function.
   - Added the toggle_button GUI widget
      - Added a default toggle button style as well as check box and 
        radio button styles.
   - Added a single click event to list_box
   - Added a save_file_box() and open_existing_file_box() function.  

Non-Backwards Compatible Changes:
   - Changed the check_box and radio_button widgets to be specializations of 
     the new toggle_button object.  This is a nearly backwards compatible 
     change except that the events registered to check_box and radio_button 
     clicks must now take the form void event(toggle_button&) or 
     void event(void) instead of the previous void event(check_box&) and 
     void event(radio_button&).
   - Removed the is_mouse_over bool from the button_style::draw_button()
     function.

Bug fixes:
   - Fixed a compiler error in mingw.
   - Changed the preprocessor checks for the wchar_t overload of 
     is_built_in_scalar_type so that it works properly with visual studio.

Other:
   - Added a Bayesian Network GUI that allows you to create a network
     and serialize it to disk.

New Features:
   - GUI Related
      - Added the scrollable_region widget
      - Added the text_grid widget
      - Added an event to the text_field so you can tell when the
        user modifies it.
      - Added the fit_to_contents() function to the tabbed_display 
        widget.
   - Bayesian Network Related
      - Added the node_first_parent_assignment(), node_next_parent_assignment(), 
        and node_cpt_filled_out() functions.

Non-Backwards Compatible Changes:
   - Reverted the change in 17.0 where I made drawable::lastx and 
     drawable::lasty not match the current location of the mouse inside
     the on_mouse_move() event.  I changed this back to how it was before,
     so now lastx and lasty represent the most current record of where
     the mouse is in *all* events.
   - Changed the functions that control text color in the label and text_field
     widgets to use rgb_pixel objects.  Also added a function to set the
     background color of a text_field.

Bug fixes:
   - Fixed a bug in the bayesian_network_join_tree object that caused it to
     compute incorrect results for some networks.
   - GUI Related
      - Fixed a minor bug in the cursor drawing of the text_field
        gui widget.
      - Fixed a bug in the compute_cursor_rect() function.  It would return an
        incorrectly positioned rectangle for 0 length strings.
      - Changed the way wchar_t is handled in the serialize.h file.  Now
        everything should compile correctly in visual studio regardless of how
        you set the /Zc:wchar_t compiler option.
      - Fixed a bug in the menu_bar widget.  One of the members wasn't being 
        initialized when it needed to be.
      - Fixed a bug in the tabbed_display where it didn't redraw itself
        correctly after it was moved by set_pos()

Other:
   - Changed the xml parser so that it counts line numbers
     from the start of the input stream instead of from the
     root tag.
   - Changed the xml parser so that you will only get the fatal_error
     event once if it occurs. 

New Features:
   - Added a zoomable_region widget
   - Added a directed_graph_drawer widget

Non-Backwards Compatible Changes:
   - Changed the first_pixel argument of the draw_string() function 
     to be a rectangle like all the other draw functions now use.

Bug fixes:
   - Fixed a bug in the tooltip widget that was triggered when calling
     its member functions without calling set_tooltip_text().  This also 
     fixed a bug in the button object that triggered when calling some button
     functions that referenced the tooltip widget.
   - Fixed a problem in the draw_circle and draw_solid_circle functions.
     They didn't draw themselves quite correctly in all cases.

Other:

New Features:
  - Added a png_loader object
  - GUI related
    - Added a popup_menu widget
    - Added a menu_bar widget
    - Added a tooltip widget
    - Added a user selectable style to the gui button.
    - Added the draw_rounded_rectangle() and fill_gradient_rounded() functions
    - Added the mouse_over_event object to the base_widgets and made the 
      button_action inherit from it.
    - Added the drawable::next_free_user_event_number() function 
  - matrix and geometry:
    - Added a size() function to matrix_exp and matrix_ref objects.
    - Added a class that represents 2D points 
    - Added the following matrix functions:
      - squared(), cubed(), get_rect(), a subm() function that takes 
        rectangles, and normalize() 
    - Added the following rectangle functions:
      - area(), centered_rect(), translate_rect(), move_rect(), resize_rect(),
        resize_rect_height(), resize_rect_width(), and nearest_point() 

Non-Backwards Compatible Changes:
  - Renamed atom() to array_to_matrix()
  - Moved the rectangle object from the gui_core into a new geometry folder  
    (only matters if you were directly including the rectangle file)
  - Moved the vector object into the geometry folder.  Also removed the kernel_1a
    stuff.  So there is now only one possible vector implementation.
  - Changed the default position for a rectangle to (0,0) instead of (1,1)
  - Added edge data to the directed_graph.  This breaks backwards compatibility 
    with the previous serialization format for directed_graphs.
  - GUI related:
    - Changed the base_window::on_keydown event signature so that it now
      reports more keyboard modifier keys (e.g. alt)
    - Made the functions for drawing on canvas objects take points and pixels 
      instead of just a bunch of integers.  Also changed the order of the
      arguments so that the canvas comes first, followed by the location
      to draw on, then what to draw.
    - Moved the canvas drawing functions into the gui_widgets/canvas_drawing.h 
      file.
    - Modified the drawable_window so that the drawable::lastx and drawable::lasty
      fields are updated after calls to on_mouse_move.  This way the x and y that
      go into the on_mouse_move actually tell you something.

Bug fixes:
  - Fixed a bug in the floating point serialization code.  It 
    didn't handle NaN or infinities correctly.
  - Fixed a bug in the win32 version of the gui_core component.  It was 
    possible that calling set_size(), set_pos(), or set_title() could cause 
    the program to deadlock.
  - Made the load_bmp() function more robust in the face of weirdly 
    written BMP files.
  - Modified the draw_circle() and draw_solid_circle() functions so that they
    only touch each canvas pixel once.  This avoids messing up alpha blending
    if an rgb_alpha_pixel is used.  

Other:
  - Removed the old win32 only gui code in the dlib/gui folder. 
  - Changed the default GUI font to a nicer Sans Serif font 

New Features:
   - Added another constructor to the thread_function object.  
     Now it can take proper function objects as well as normal function
     pointers.
   - Added the probabilistic_decision_function object and svm_nu_train_prob() 
     function.

Non-Backwards Compatible Changes:
   - Changed the svm train functions so that the cache_size argument
     now measures the max number of megabytes of memory to use rather
     than number of kernel matrix rows to cache.  It's default
     value is now 200MB.
   - changed the type typedef in the SVM kernel function objects to
     be named sample_type instead of type. 

Bug fixes:
   - Fixed a bug in the trim, rtrim, and ltrim functions.  They
     didn't return empty strings when the input string contained all
     trim characters.
   - Fixed a bug in the decision_function's copy constructor

Other:
   - Added an optimization to the working set selection for the svm training code.
     Now the algorithm will prefer to select indices that are in the kernel 
     matrix cache when possible.
   - Fixed a problem with the chm documentation file where many of the links
     didn't work.
   - Made the support vector functions capable of operating with floats, doubles,
     and long doubles instead of just the double type.

New Features:
   - Added aversion of the draw_line() function for images.
   - Added the atom(), rowm(), colm(), and subm() matrix functions.
   - Added some push/pop_back() functions to the array object that are similar
     to the ones in the std::vector.
   - Added the std_vector_c class that wraps std::vector and checks its 
     function's preconditions.
   - Added the polynomial_kernel object for use with the svm algorithm.

Non-Backwards Compatible Changes:
   - Changed the svm_nu_cross_validate() function to return a vector
     of both the +1 and -1 cross validation accuracies.

Bug fixes:
   - Fixed a bug in the list_box that caused it to not hide itself properly
     when told to do so.
   - Fixed canvas::fill() gui function so that it should work right
     on 64 bit platforms.

Other:

New Features:
   - Added memory manager support to the matrix object.

Non-Backwards Compatible Changes:
   - Made the assign_pixel() function saturate grayscale values bigger
     than the target pixel type can handle.  Previously it would just 
     truncate the numbers.
   - Removed rand_kernel_1 and rand_kernel_2 because they gave very 
     inferior results compared to rand_kernel_3.  I then renamed
     rand_kernel_3 to rand_kernel_1.
   - Renamed rand::get_random_number() to get_random_8bit_number() and also 
     added a get_random_16bit_number() and get_random_32bit_number()
   - Added a checksum to compress_stream_kernel_1 and kernel_2.  This 
     breaks backwards compatibility with the previous versions.  That is,
     the new implementations will complain that decompression fails if 
     you give them data compressed with the old non-checksum version of
     the compression routines. 
   - Removed the width() and height() functions from the array2d object.  
     Now only the equivalent nc() and nr() member functions remain.
   - Changed array2d::set_size(width,height) to set_size(num_rows, num_cols).  
     That is, I switched the order of the two arguments to this function.
     The reason for doing this is to make it have the same form as the
     set_size() member of the matrix object.  This way the usage of the
     set_size() member for these two very similar data structures is 
     the same.  Hopefully this will reduce confusion rather than 
     make things worse.

Bug fixes:
   - Fixed a bug in the image_widget.  It didn't repaint the screen
     all the way if you gave it a smaller image to display.
   - Fixed a bug in the cat() function that caused the state of the queue
     to be broken if you called cat with an empty queue.
   - Made the queue_sort_1 use a better sorting algorithm.  In particular, it 
     will not sort slowly for nearly sorted data.
   - Fixed a bug in the queue_kernel_2 object that caused it to not work 
     correctly with the non-default memory managers.

Other:
   - Added example code for the member_function_pointer as well as the matrix 
     object.
   - Added some more regression tests and made some of the longer running 
     ones execute a lot quicker.
   - Made the unit test suite easier to use.  Now tests just throw an exception 
     to indicate an error rather than returning an error code.
   - Added an example program for the multi-layer perceptron neural network.

New Features:
   - Added the is_signed_type and is_unsigned_type templates
   - Image Processing stuff
      - Added the assign_all_pixels() function
      - Added the assign_border_pixels() function
      - Added the assign_pixel_intensity() function
      - Added the auto_threshold_image() function
      - Added the binary_union() function
      - Added the edge_orientation() function
      - Added the get_histogram() function
      - Added the get_pixel_intensity() function
      - Added the hysteresis_threshold() function
      - Added the sobel_edge_detector() function
      - Added the suppress_non_maximum_edges() function
      - Added the zero_border_pixels() function
      - Changed the pixel_traits structure so that it can support 8, 16, and 32 
        bit grayscale pixels. 

Non-Backwards Compatible Changes:
   - Added more fields to the pixel_traits template so if you had defined your 
     own pixel types you will need to update them.

Bug fixes:
   - Fixed some compiler errors in Visual Studio 2008

Other:
   - Generally tried to clean up the documentation and code in this release

New Features:
   - Added the randomize_samples() function
   - Added the set_main_font() and main_font() functions to the drawable object.
     So now the drawable widgets can use a user provided font.  

Non-Backwards Compatible Changes:
   - Made the named_rectangle object a little easier to use.  It now won't
     let you size it so small that it doesn't display its entire name.

Bug fixes:
   - Fixed a bug in the svm_nu_train() function that caused a crash with 
     some inputs.
   - Fixed a compile time error that occurred when compiling the bayesian
     network code in Mac OS X.
   - Fixed a bug in the compute_cursor_pos() function where it would 
     return the incorrect value.
     
Other:
   - Added an example showing how to use the svm functions.

New Features:
   - Added the left_substr() and right_substr() functions
   - Added the zero_extend_cast() function
   - Added the unsigned_type template
   - Added the uint8 typedef
   - Bayesian Network related
      - Added the assignment object
      - Added the bayes_node object
      - Added the joint_probability_table object
      - Added the conditional_probability_table object
      - Added the bayesian_network_gibbs_sampler object
        This object implements an algorithm that performs approximate inference
        in a Bayesian Network.
      - Added the bayesian_network_join_tree object
        This object implements an algorithm that performs exact inference
        in a Bayesian Network.
   - Set related
      - Added the set_intersection_size() function
      - Added the set_union() function
      - Added the set_intersection() function
      - Added the set_difference() function
   - Graph related
      - Added the graph object
      - Added the is_graph template
      - Added the is_directed_graph template
      - Added the create_moral_graph() function
      - Added the triangulate_graph_and_find_cliques() function
      - Added the graph_contains_length_one_cycle() function
      - Added the find_connected_nodes() function
      - Added the graph_is_connected() function
      - Added the is_clique() function
      - Added the is_maximal_clique() function
      - Added the copy_graph_structure() function
      - Added the create_join_tree() function
      - Added the is_join_tree() function
      - Added the edge() function
   - GUI related
      - Added the base_window::get_display_size() function
      - Added message_box_blocking()
      - Added the bdf_font object which is capable of loading BDF font files into
        the font object used by the gui_widgets
   - Better Unicode support
      - Added the basic_utf8_ifstream: An input stream that can read UTF-8 files
      - Added serialization support for wchar_t and std::wstring
      - Added the is_combining_char() function
      - Added the convert_utf8_to_utf32() function
      - Modified most of the string manipulation functions in dlib/string.h
        to work with any kind of character type
      - The gui widgets' font object now works with Unicode text (i.e. wchar_t 
        and unichar strings) as well as with normal char data.

Non-Backwards Compatible Changes:
   - The dlib/all_console.cpp and dlib/all_gui.cpp files have been deprecated
     in favor of a new file.  Now to build with dlib you simply add 
     dlib/all/source.cpp to your project regardless of what type of project
     you are building.
   - The GUI program entry point, winmain(), has been removed.  You can now use 
     the normal main() entry point or some other non-standard entry point 
     provided by your compiler.
   - Renamed directed_graph::node::item to directed_graph::node::data

Bug fixes:
   - Fixed some build issues in gcc 4.2 involving some uses of the std_allocator
   - Fixed some build issues in Visual Studio involving the dir_nav component
     and building with NO_MAKEFILE #defined.
   - Moved the #define that disables the old WinSock API into the sockets cpp 
     file.  This should avoid conflicts with people who are using the old WinSock
     API.
   - Changed the tuple template slightly to avoid a bug in Visual Studio 7.1 
     that caused a compile time error in some instances.
     
Other:

New Features:
   - Added a destroy() function to the map, set, hash_map, and hash_set objects.
   - Added the tuple object
   - Added an overload of connect() that has a timeout
   - Added rand_kernel_3 as a random number generator that uses the Mersenne Twister 
     algorithm.
   - Added the directed_graph object
   - Added the graph_contains_undirected_cycle() and graph_contains_directed_cycle()
     functions.
   - Added the std_allocator object.   It is a STL style allocator that can use 
     the dlib memory manager objects.
   - std::string manipulation functions:
      - Added the cast_to_string() function.
      - Added the tolower() function
      - Added the toupper() function
      - Added the ltrim() function
      - Added the rtrim() function
      - Added the trim() function
      - Added the lpad() function  
      - Added the rpad() function  
      - Added the pad() function  

Non-Backwards Compatible Changes:
   - Changed the default logging level from LNONE to LERROR
   - Renamed the ASSERT macro to DLIB_ASSERT and CASSERT to DLIB_CASSERT.  
     This rename avoids a conflict with a macro inside MFC.
   - Changed the logger so that settings are inherited when a new logger
     is instantiated rather than just having the new logger use the 
     default settings.
   - Removed the logger::clear() function since it no longer really
     makes sense given the above change.
   - Removed the get_main_thread_id() function and replaced it with the
     is_dlib_thread() function.

Bug fixes:
   - Pushed some things into cpp files because doing so avoids build and/or 
     runtime errors on some platforms.   

Other:
   - Changed the string_cast() function so that it will recognize the words true 
     and false as boolean values.  Also improved the error message inside the 
     string_cast_error exception object.

New Features:
   - Added the covariance() function
   - Added the rgb_alpha_pixel pixel type and modified all relevant functions to
     support it.

Non-Backwards Compatible Changes:
  - The GUI changes that are non-backwards compatible:
       - The alpha parameter is now an unsigned char instead of unsigned int
         and its range is now 0 to 255 instead of 0 to 256.
       - The image_widget no longer has any member functions dealing with
         alpha values.  If you want to use alpha blending you just give it an
         image that has an alpha channel.  The same goes for draw_image().
  - There are now more fields in the pixel_traits template.  So if you were 
    defining your own pixels before you will need to update your pixel_traits 
    specializations.

Bug fixes:
  - Made some functions non-inline and put some things on the stack 
    instead of heap.  Doing this avoids some problems with certain 
    kinds of builds in visual studio.

Other:
  - Modified the message_box() function so that it is safe to call end_program() 
    from within its callback event.

New Features:
   - Modified the GUI drawing functions to take an alpha argument to allow
     alpha blending.
   - Added the svm_nu_cross_validate() function to perform k-fold
     cross validation using the svm_nu_train() function.
   - Added the boost enable_if templates
   - Added the rand_float extension to the rand object.
   - New matrix features:
      - Added the pinv() function
      - Changed round_zeros() to use the machine epsilon instead of 1e-6 as 
        its default epsilon.
      - Modified the matrix object so that you can declare them with 
        a static dimension and a dynamic dimension.  E.g. matrix<float,0,10>
        is now legal and declares a matrix with a fixed number of columns(10) 
        and a variable number of rows.
      - Added the equal() function to compare two matrices of floating
        point numbers for near equality.
      - Changed the matrix so that operator(long) works for both
        column vectors and now also for row vectors.
      - Added a set_size() and constructor that takes a single long for use in
        sizing row and column vectors.
      - Added the scale_columns() function

Non-Backwards Compatible Changes:

Bug fixes:
   - Fixed an error in svm_nu_train() where it would incorrectly
     complain of incorrect nu values for some datasets.
   - Added a missing std:: qualifier at two points in the dlib/vector code that
     could cause a compiler error in some instances. 

Other:
   - Added a term index to the documentation.

New Features:
   - Added a nu support vector classifier training function.
   - Added a multilayer neural network object.
   - Added the "destructive aliasing" checks into the matrix code.  Now temporary 
     matrices are only created during assignment if the right hand side aliases 
     the left hand side in a destructive way.  This removes many of the previous
     uses of temporary matrices.
   - Made the sum() matrix function be able to sum matrices of matrices
   - New matrix functions:
      - acos(), asin(), atan(), ceil(), cos(), cosh(), exp(), floor(), log(), 
        log10(), mean(), norm(), pow(), reciprocal(), round_zeros(), sin(),
        sinh(), sqrt(), tan(), tanh(), variance(), and more overloads of 
        uniform_matrix().

Non-Backwards Compatible Changes:

Bug fixes:
   - Added missing nr() and nc() functions to the uniform_matrix() and 
     identity_matrix() functions.
   - Forgot to add a destructor for the dynamically sized matrix resulting in a 
     memory leak.  This is now fixed.
   - Fixed various potential compile time errors

Other:

New Features:
   - Added a copy of the boost noncopyable base class.
   - added some smart pointers:
      - added shared_ptr
      - added weak_ptr
      - added scoped_ptr

Non-Backwards Compatible Changes:

Bug fixes:

Other:
   - Cleaned up the assert code and removed the need for the dlib/error.ccp file
   - Made the matrix take better advantage of the compile time sized
     dimensions when it can.  

New Features:
   - Made it so that command line options have a default conversion to bool
     and the bool tells you if they are on the command line or not.
   - Added an implicit conversion to a scalar to the matrix object 
     when it is of dimension 1x1.
   - Added the thread_function object
   - Added a function to compute the singular value decomposition of a matrix.

Non-Backwards Compatible Changes:
   - Added two new arguments to the on_request() function.  They allow you to
     see what HTTP headers the client sends you and to control which ones 
     you send back.

Bug fixes:

Other:

New Features:
  - matrix object additions:
     - Added some functions to convert between matrix and pixel objects.
     - Added the clamp() function that operates on matrix objects.
     - Added the sigmoid function.
     - Made the matrix object capable of being sized at runtime in addition 
       to its original compile time static sizing capability.
     - Added 3 and 4 argument versions of pointwise_multiply()
     - Added the +=, -=, *= and /= operators to the matrix object.

Non-Backwards Compatible Changes:

Bug fixes:
  - Fixed the line numbering in the color pretty printer.  Wasn't being
    done correctly.
  - Fixed a bug in the matrix round() function.
  - Fixed some miscellaneous compile time errors
  - Fixed a bug in the matrix removerc() function.
  - Added some missing checks to catch invalid negative index inputs to
    matrix objects.
  - Fixed a bug in the matrix inv() function.  It could sometimes 
    segfault if used on certain singular matrices 

Other:
  - string_cast() can now convert hex strings to integers
  - You can now say myarray2d.set_size(0,0) and have it do what
    you would naturally expect.
  - Added some #pragma statements that tell visual studio
    to link the right system libraries automatically.  
    So now you don't have to add these things in the 
    project settings anymore. 

New Features:
  - Added the set_all_logging_levels(), set_all_logging_output_streams() 
    functions
  - Added the configure_loggers_from_file() function which allows you to
    easily configure all logger objects using a textual configuration
    file.

Non-Backwards Compatible Changes:

Bug fixes:
   - Added a workaround into the code that avoids a potential compilation
     error on Mac OS X systems.

Other:

New Features:

Non-Backwards Compatible Changes:

Bug fixes:
   - Fixed a bug in the POSIX version of the hostname_to_ip() function. It was
     screwy if you asked for more than the first IP address (the same address
     might be returned more than once).
   - Fixed a bug in the pipe object's timeout functions.  The timeouts weren't
     working correctly.

Other:

New Features:
   - Added the wait_for_num_blocked_dequeues(), enable_enqueue(), 
     disable_enqueue(), and is_enqueue_enabled() functions to the pipe object.
   - The pipe object can now be used with a zero length buffer.

Non-Backwards Compatible Changes:
   - There is no longer a pipe::kernel_1a_c typedef since the pipe 
     no longer has any requirements to check (due to the change of allowing
     zero length buffer sizes)

Bug fixes:

Other:
   - Made the ASSERT and CASSERT macros call dlib_assert_breakpoint() when they 
     fail.  This way you can easily set a breakpoint on them in a debugging
     tool by breaking on calls to this new function.
   - Fixed some typos and unclear parts of the pipe spec. 

New Features:
   - Added a thread safe version of the config_reader object (in the form of an 
     extension to the config_reader)
   - Added the wait_until_empty() function to the pipe object.

Non-Backwards Compatible Changes:
   - Removed the connection::close() and listener::close() functions.  They have
     been replaced by destructors.  To upgrade old code all you have to do is 
     replace statements of the form "object->close();" with "delete object;".  
     Both statements do exactly the same thing.  However, for connection objects, 
     you should probably be using the close_gracefully() function instead.

Bug fixes:
   - Removed a potential compile time error in the dng image format handling code.
   - Fixed a bug in the bigint object.  The destructor was using "delete" 
     when it should have been using "delete []"
   - Fixed a resource leak in the POSIX version hostname_to_ip()
   - Fixed a significant memory leak in memory_manager_kernel_1
   - Fixed a memory leak that could occur in memory_manager_kernel_2
     and memory_manager_kernel_3 when the constructor for the object
     being constructed threw an exception.
   - Added a missing delete statement to plug a memory leak
     in the md5 computation code.
   - Fixed an uninitialized variable warning from valgrind 
     (in lz77_buffer/lz77_buffer_kernel_2.h).  I think this could
     also potentially result in an error when decoding data but I'm not totally
     sure.  But either way it is fixed now.
   - Changed a call to memcpy to memmove in the sockstreambuf_kernel_2 
     implementation since the copy could potentially be of overlapped memory.  

Other:
   - Changed the connection::read() and connection::write() functions to take
     longs instead of unsigned longs as the buffer sizes.  They also now 
     return longs instead of ints.  This should be a backwards compatible change.
   - Started using the valgrind tool to check for memory errors in the project and
     found a few things.  Very nice tool :)

New Features:
   - Added the multithreaded_object extension to the threads API
   - Added the load_dng() and save_dng() functions which can load and store
     the DNG lossless compressed image format (which I just made up).

Non-Backwards Compatible Changes:
   - Changed the serialization format for bool to use a 1 byte code rather than 2 
     bytes.  So this breaks compatibility with the old format.

Bug fixes:
   - The serialization for bool didn't always work right.  This is now fixed. 

Other:

New Features:
   - New faster version of the bigint object (bigint_kernel_2) that uses
     the Fast Fourier Transform to perform multiplications.
   - The base_window can now be an "undecorated" window.  This new type is suitable
     for making things like popup menus and the like.
   - Added the on_focus_lost() event to the base_window object
   - Added the on_focus_gained() event to the base_window object
   - Added the on_window_moved() event to the base_window object
   - Added the get_pos() function to the base_window object
   - Updated the gui_widgets's drawable interface stuff to support the three
     new event types and the new window type.
   - Added the drawable::draw_rectangle() function
   - Added serialization support for std::complex.
   - Added the assign_image() function

Non-Backwards Compatible Changes:
   - Removed the color arguments from the drawable_window object's constructor and
     added a new boolean argument (if it is an undecorated window or not).  This
     probably won't break any code but if it does you should get a compiler error
     about it.
   - Made it so you must disable the events in the destructor for your
     drawable gui widgets.  Doing so avoids potential race conditions when
     destructing drawable objects.
   - Made it so that you are required to call close_window() in a window object's 
     destructor.  This avoids a potential race condition.

Bug fixes:
   - Added a workaround for a bug in MinGW that caused the regression test suite
     to crash when compiled with -O3.
   - Fixed a potential bug in the X Windows version of the gui_core component. 
      Added an extra XFlush() to end_program() because without it a 
      program can crash when calling end_program() in certain instances.
   - The spec for the pipe object said that objects you enqueue into it
     had an "initial value for their type" after the function completes. This
     is incorrect, they are swapped into the pipe so they have an undefined
     value afterwards.  I fixed the spec for the pipe to say this.
   - Fixed a bug in the font rendering functions in the gui_widgets
     component.  It could cause a segmentation fault sometimes.
   - Fixed some potential deadlocks in the windows version of the gui_core
     component.
   - Fixed a bug in the rsignaler object.  When you called wait() or 
     wait_or_timeout() it only unlocked the associated rmutex once (it could be 
     locked more than once and thus might cause a deadlock since the thread
     calling wait() wouldn't actually unlock the mutex in this case).  
   - Fixed the initialize_array() function in memory_manger_kernel_3 to be 
     exception safe.  Previously if an exception occurred while creating
     an array then a resource leak was created.

Other:
   - Changed the package format for the library somewhat.  The examples are now 
     located in their own top level folder.  Additionally, the HTML version of the 
     documentation also comes in the same archive as the source rather than in a 
     separate download.
   - Started using major and minor version numbers rather than just major ones.

New Features:
   - Added operator<< and operator>> iostream operators to the vector object. 

Non-Backwards Compatible Changes:
   - Changed the xml_parser's document_handler interface:  
     made empty element tags (<like_this/>) trigger the end_element() callback
     and removed the is_empty bool from start_element().

Bug fixes:
   - Fixed a potential race condition between the destruction of the thread pool
     and the "program ending handlers" stuff.

Other:
   - Made the xml parser more robust to different types of new line characters.
   - Modified the source slightly so that it works with mingw.

New Features:
   - The config_reader is now enumerable.  
   - Added the image_widget gui object.
   - Added nr() and nc() to the array2d object.
   - Added the shutdown_connection() function to the iostream extension
     to the server object.
   - Added the timer_kernel_2 implementation which is a version of the timer object
     that is more efficient in its allocation of threads.
   - Added the timeout object.
   - There is now a CMakeLists.txt file located in the dlib folder.  See
     dlib/examples/CMakeLists.txt and dlib/test/CMakeLists.txt for examples 
     that use CMake to build projects using this library.
   - Added the register_program_ending_handler() function to the threading API.

Non-Backwards Compatible Changes:
   - Removed the config_reader::get_blocks() function.  Use the
     new enumerable interface for the config_reader instead.
   - The array2d object now uses longs instead of unsigned longs to report
     its dimensions and access its elements.  
   - Added a uint64 to the on_connect() callback in the iostream 
     extension to the server object.
   - timer::set_delay_time() now throws and timer::start() now may throw 
     std::bad_alloc.

Bug fixes:
   - Fixed a bug in end_program().  In X Windows it might not cause the
     program to end right away if it was called from outside the event
     handling thread.
   - Fixed a bug in the implementation of the timeout part of the
     close_gracefully() function.

Other:
   - The library now works on HP-UX
   - The regression test suite now has command line arguments that
     enable tests to send debug messages to a file.

New Features:
   - The http server extension now supports the POST HTTP method.
   - The attribute list object in the xml_parser is now enumerable.
   - Added the threaded object extension
   - Added the uintn.h file which defines fixed sized unsigned integral types.

Non-Backwards Compatible Changes:
   - Renamed the on_get() callback in the http extension to the server object to
     on_request()
   - Removed the network byte order functions from the sockets api.  (They are still 
     really there though since they come from actual OS header files.  But 
     officially they have been replaced by the byte_orderer component).
   - Renamed dlib/uint64.h to dlib/uintn.h

Bug fixes:

Other:
   - The command line parser will now let you declare long named options with - 
     characters in them.
   - Made it so you can use the COMPILE_TIME_ASSERT macros anywhere rather than
     just inside functions.

New Features:
   - For dlib::matrix
      - Added the tmp() function
      - Added optimized specializations of inv() and det() for 1x1, 2x2, 3x3 and 
        4x4 matrices.
      - Added the removerc() function 
   - Sockets related
      - Added the connect() function
      - Added the is_ip_address() function.
      - Added the close_gracefully() function
      - Added the iostream extension to the server object.
      - Added the http extension to the server object.

Non-Backwards Compatible Changes:
   - Changed the cpp_tokenizer to not convert characters to their html form.

Bug fixes:
   - Removed some potential compile time errors.  See the change log for details.

Other:
   - Improved the web site
   - Added some more example code
   - Added more colors to cpp_pretty_printer_kernel_1.

New Features:
   - std::map is now serializable
   - Added the matrix object and a bunch of supporting code.  
   - Added the list_box graphical widget
   - Added the fill_rect_with_vertical_gradient() function to the
     drawable interfaces list of drawing helpers.
   - Added the open_file_box() function which provides a simple file chooser.

Non-Backwards Compatible Changes:

Bug fixes:
   - Made timestamper::get_timestamp() be a const function like it should.  Fixes 
     some compile errors.
   - Fixed a bug in the font::draw_string() function.  It didn't redraw
     multi-line strings right.
   - Fixed a bug in the scroll_bar object that would cause a compile
     error if you tried to call its width() function.
   - Fixed a bug in the array_kernel_1 object.  It would cause a segmentation fault 
     when used sometimes.

Other:

New Features:
   - Added the following image transformation functions:
      - Added the equalize_histogram() function
      - Added the spatially_filter_image() function
      - Added the threshold_image() function
      - Added the binary_dilation() function
      - Added the binary_erosion() function
      - Added the binary_open() function
      - Added the binary_close() function
      - Added the binary_intersection() function
      - Added the binary_difference() function
      - Added the binary_complement() function
   - Added the clear(), load_from() and default constructor back into the 
     config_reader. 
   - Made the member_function_pointer copyable and also added operator== and != 
     to it.

Non-Backwards Compatible Changes:
   - Made the vector object templated so you can use types other than double with it.
     But now you 
     have to specify what type you want to use which is slightly different.
   - The asc_pair_remover and asc_remover abstract classes now take a third template 
     argument.  I highly doubt this effects any code outside the library but it is 
     possible.
 
Bug fixes:
   - Fixed a bug in the base_window::set_size() function.  If you specified a size
     of (0,0) it caused your program to error out.  This has now been fixed.
   - Fixed a bug in the scroll_bar widget.
   - Fixed a bug in save_bmp().  For some image sizes it would output a goofy 
     looking skewed image.

Other:
   - Switched everything that used to call operator< directly to instead use 
     std::less or to take a template argument that provides a compare functor that
     defaults to std::less.

New Features:
   - Added the assign_pixel() function
   - Added the hsi pixel type 
   - Added the save_bmp() function
   - Added the static_switch template

Non-Backwards Compatible Changes:
   - Changed how the config_reader works.  It now has a more powerful syntax and 
     improved interface.  Converting any old code to use the new version should be 
     simple since the new file syntax is very nearly backwards compatible with the 
     old syntax.  (i.e. You probably won't have to change anything about your 
     config files)
   - Renamed the dlib/image_loader.h file to dlib/image_io.h since it now includes
     the image saver stuff.
   - Renamed the pixel struct to rgb_pixel
   - Renamed pixel_traits::color to pixel_traits::rgb
   - Renamed pixel_traits::scalar to pixel_traits::grayscale
 
Bug fixes:
   - Fixed a bug in the load_bmp() function.  It would load 24bit bmp files 
     incorrectly.
   - Changed the logger so that it won't deadlock if you write something similar to
     my_log << LINFO << function_that_also_logs();.  Although this is a
     dumb thing to do.  But even so, it shouldn't deadlock.
   - Fixed a potential linking problem with the vector object.

Other:
   - I decided I'm not going to support Borland v5.5.1 anymore.  There are just too 
     many bugs in this compiler.  It is very old at this point so I don't see this 
     being a big deal. 
   - Made the drawable::draw_image() and load_bmp() functions able to handle images
     of any type of pixel.
   - Pulled the imaging, algorithmic and metaprogramming stuff out of the 
     miscellaneous section of the web page and gave them all their own sections.

New Features:
   - Added a logger header that prints the date and time.
   - Added the LTRACE logging level
   - Added a buffered implementation of sockstreambuf. 

Non-Backwards Compatible Changes:
   - Changed the specs to say that sockstreambuf may be buffered.  
     sockstreambuf_kernel_1 is still just as it always has been though.  So all old 
     code will still work just as it always has.  But all the same, the specs have 
     been changed and now allow for an implementation that is not 100% backwards 
     compatible.
   - rand_kernel_2 now emits a different string of random numbers.  
 
Bug fixes:
   - Changed the logger object's implementation to not try to register
     a thread end handler for the main program thread anymore.  This was 
     technically a bug according to the spec but it actually did end up
     working the way it was supposed to.  But even so, it shouldn't have 
     been doing that.
   - Changed binary_search_tree_kernel_1 so that it avoids a bug in the version of
     gcc on SuSE Enterprise Linux 9.
   - Fixed a bug in the rand_kernel_2 implementation.  It wasn't giving good 
     random numbers. 

Other:
   - Modified the code so that you don't get any warnings when -Wall is used with 
     GCC.

New Features:
   - Added the ASSERT_ARE_SAME_TYPE macro
   - Added the is_same_type template
   - Added the get_main_thread_id() function to the threading API
   - Added the thread_specific_data extension to the threading API
   - Added the logger object.
   - Added the auto_unlock object to the threading API.

Non-Backwards Compatible Changes:
 
Bug fixes:

Other:
   - Added an example that is specifically about using threads
   - Added two examples about using the logger object

New Features:
   - Added the memory_manager_stateless object and two implementations of it. 
   - Added the MACOSX macro to dlib/platform.h
   - Added a templated version of create_new_thread() that allow you to start
     a thread with a member function.
   - Added the register_thread_end_handler() function to the threading kernel API. 
   - Added memory_manager_kernel_3

Non-Backwards Compatible Changes:
   - Changed the meaning of the memory_manager_global::get_number_of_allocations() 
     function because the previous definition of it didn't really make sense for 
     this object.
   - Changed the threading API to wait for all outstanding threads to terminate 
     before allowing the program to end.  It used to just let the OS trash those 
     threads when main() ended but this isn't a safe or portable thing to do.  I 
     used to assume the user would make sure all their threads were done and had 
     cleaned up whatever they were doing before main() ended but this is too much 
     of a burden on the end user.  So now the library will wait for your threads to
     end.  You still need to have some way of telling them it is time to stop though.
 
Bug fixes:
   - Fixed a minor bug in dlib/timer/timer_kernel_1.h.  Its implementation was 
     slightly off according to the specification of a timer object.

Other:
   - The byte_order object is now capable of flipping entire arrays.
   - Made it so that the ENABLE_ASSERTS macro is defined whenever ASSERT is
     on.
   - Made the array container use the memory managers.

New Features:
   - Added functions to explicitly convert to/from little and big endian to the 
     byte_order object.
   - Added the allocate_array() and deallocate_array() functions to the 
     memory_manager.
   - Created the memory_manager_global object
   - Added the remove_last_in_order(), position_enumerator() and 
     remove_current_element() functions to the binary_search_tree object.

Non-Backwards Compatible Changes:
   - I put an #error directive in the old GUI component to notify anyone
     trying to use it that it is deprecated.  I will be removing it from the
     library in a few months.
   - Switched the reference_counter object back to not using the memory_manager.
     I realized it isn't safe for this object to use the memory_manager since 
     it could result in memory_managers freeing each other's allocations.
   - I redefined the pixel_traits template.  It is now a lot simpler and more 
     convenient.
 
Bug fixes:
   - Fixed a minor bug in dlib/rand/rand_kernel_2.cpp

Other:
   - Added some more compile time checks to the byte_orderer object.
   - Changed some includes and preprocessor macros around a little so now
     everything but the GUI stuff compiles in mac OS X.
   - Added inclusion guards to all the .cpp files
   - Added the all_gui.cpp and all_console.cpp files.  They 
     include all the .cpp files you need to make gui and
     console applications respectively into one file.
   - Made more containers use the memory_manager.

New Features:
   - Added the enqueue_or_timeout() and dequeue_or_timeout() functions
     to the pipe object.
   - Gave the mouse_tracker the ability to display the mouse position
     relative to a user selected point.
   - Added the message_box() function to the gui_widgets component.
   - Gave the label widget the ability to draw newlines in strings.
   - added the close_window() and is_closed() methods to the base_window
     object.
   - Added the rsignaler extension to the threading API.
   - You can now control the thread pool timeout time by setting the
     DLIB_THREAD_POOL_TIMEOUT #define.
   - Added the get_from_clipboard() and put_on_clipboard() functions
     to the gui_core component.
   - Added the stop_and_wait() function to the timer object.
   - Added the trigger_user_event() function and on_user_event() event
     to the base_window object.  This new event is also forwarded 
     to drawable interfaces inside the receiving window.
   - Added the wrap_around() function to the named_rectangle widget.
   - Added the top(), left(), right(), bottom(), width() and height() 
     functions to the drawable interface. 

Non-Backwards Compatible Changes:
   - Made the radio_button and check_box widgets pass references to themselves
     when they call their click handlers.  
   - Switched the sync_extension to use the rmutex and rsignaler objects
     rather than the normal non-reentrant ones. ( Chances are that old
     code that used this will still compile fine anyway. )
   - Changed the return type of rand::get_random_number() to be an 
     unsigned char.  I also changed both the implementations of 
     rand because they weren't very good at all. 
   - Changed the functions related to drawing strings in the font class.
   - Changed the drawable's rectangle to default to being empty
     rather than being a single point.  Most code should not notice
     the difference.
 
Bug fixes:
   - The event handlers in gui_widgets/drawable.h were private.  They
     should be protected.  This is now fixed.
   - Fixed a bug in the way the scroll_bar was drawn when it was
     the HORIZONTAL type.
   - Changed how the thread pool destructs itself when the program 
     is ending.  Previously it was possible to get an error on
     NetBSD when the program was ending.  This is now fixed.
   - The functions related to setting the jump size in the scroll_bar
     widget were private.  They are now public.
   - There was a bug in the MS Windows version of the gui_core component
     where the members of the base_window would not work if called from
     within the on_window_close() event.  This has now been fixed.
   - Made the set_pos() function work right for the mouse_tracker widget.
   - Fixed a bug in the base64 object where the string "" could potentially 
     be decoded incorrectly.
   - Made the global swap function for crc32_kernel_1 inline.  This fixes a
     potential linker error.
   - Fixed some potential deadlocking that could occur while using the
     gui widgets.  

Other:
   - I moved all the regression tests into the dlib/test folder and
     made a nice driver program to run all of them.  
   - I have been using the sourceforge compile farms to test the library
     on various platforms.  It now works for Solaris and some of the BSDs 
     in addition to Linux and Windows.

New Features:
   - Added the array_expand extension to the array object.
   - Added the cmd_line_parser_check extension to the command line parser.
   - Added the pipe object.
   - All applicable container classes now use the memory_manager component for
     their memory allocation.
   - New implementations of the memory_manager object. 
   - Added the copy_functor class.

Non-Backwards Compatible Changes:
   - Moved the wrap_string, narrow, and string_cast functions
     to a new file.  You now have to include dlib/string.h to get
     them.  (This makes a bunch of other things work right in gcc 2.95)
   - Renamed the _L macro to _dT
   - Removed the scopy class
   - Simplified the interface to the memory manager.  It is basically the same
     though.  
   - Removed the max_size() methods from the hash_table and binary_search_tree 
     objects.
   - Removed the T_is_POD template arguments from the hash_table and 
     binary_search_tree objects.
   - Simplified the template arguments to all checking components and extensions. 
     They now take the class they are to extend or check as their only template
     argument.  This only affects you if you haven't been using the kernel_nx 
     typedefs.
 
Bug fixes:

Other:
   - I changed a few things around and now a majority of the library
     again compiles under gcc 2.95.  But some things don't and I currently
     don't plan on making them work because it involves hackish workarounds
     of 2.95's bugs. 
   - Changed the compress_stream_kernel_1 object so that it will detect data
     corruptions better.  This change will prevent it from correctly decompressing
     data that was compressed with a previous version and has an uncompressed size
     greater than about 20,000 bytes.
   - There is a new cpp file you need to compile:  dlib/error.cpp
   - Moved all the regression testing stuff into the dlib/test folder and made
     a nicer test driver to run them.

New Features:
    - Created the byte_orderer object.
    - Created the mouse_tracker gui widget.
    - The sliding_buffer object is now enumerable and serializable.
    - Added the get_filesystem_roots() function to the dir_nar component.
    - Added the create_directory() function to the misc_api component.

Non-Backwards Compatible Changes:
    - The ASSERT macro is now only enabled if DEBUG or ENABLE_ASSERTS
      is defined.
 
Bug fixes:
    - Fixed a minor bug in the cmd_line_parser object.  If you gave
      an option such as --option=arg when option didn't take any 
      arguments it could hang your program.
    - Fixed a bug in wait_or_timeout() in the posix version of the threading 
      api.  The time to wait for was being calculated incorrectly and could 
      result in an excessive number of spurious returns.
    - Fixed a minor bug in the on_keydown() event for windows.
      I had it set such that the shift and ctrl bools would be false
      if they were the actual keys being pressed.  This isn't what the
      specs say should happen but I had a comment in the windows code
      that made it clear that I did it on purpose.  Go figure :)
      This is now fixed. 

Other:
    - Improved the cpp_tokenizer object's ability to recognize numerical 
      constants.
    - Improved the text_field gui widget.
    - There are now two assert macros.  One called ASSERT
      and another CASSERT.  They both do the same thing but ASSERT
      is only enabled when DEBUG or ENABLE_ASSERTS is defined.
      All the old ASSERT statements were changed to CASSERT statements.

New Features:
    - Added array_kernel_2 which is a simple layer on top of a C array.
    - Added the tabbed_display GUI widget 
    - Added the widget_group GUI widget 
    - Added the named_rectangle GUI widget 
    - Added the pixel_traits template

Non-Backwards Compatible Changes:
    - The default maximum size for an array object is now 0 rather than 
      20,000.
Bug fixes:

Other:
    - made the cpp_pretty_printer a little better about how it handles
      C style code.  Also added support for /*!A html_anchor_name !*/ 
      style comments.

New Features:
    - Created the array2d object.
    - Created the base64 object.
    - Created the pixel struct.
    - Created the load_bmp() function which can load a BMP image file
      into an array2d object of pixels.
    - Created the drawable::draw_image() function

Non-Backwards Compatible Changes:
    - In the drawable interface I made the z order a long rather
      than unsigned long.
    - The cpp_tokenizer object now has a NUMBER token type.
    - removed the get_ prefix from functions in the cmd_line_parser 
      and cmd_line_parser_option objects.  Also changed the 
      cmd_line_parser_option::operator[] function to a normal member 
      function called argument(). 
 
Bug fixes:

Other:
    - cpp_pretty_printer now colors numeric literals a shade of yellow.

New Features:
    - Created the member_function_pointer object.
    - Created the button_action object.
    - Created the arrow_button object.
    - Created the check_box object.
    - Created the radio_button object.
    - Created the scroll_bar object.
    - More drawing functions to draw various things
      onto a canvas object.
    - Added enable/disable functions to the 
      drawable interface.
  
Non-Backwards Compatible Changes:
    - The gui widgets are no longer templated at the 
      class level.
    - The drawable object's constructor now takes a 
      bit set rather than a bunch of bools to tell it
      which events to enable.
    - I changed the names of some of the functions in the
      gui_widgets component so that they all reflected a 
      uniform naming style.
 
Bug fixes:
    - Fixed a minor bug in the cpp_tokenizer.  
    - Minor bug in the timer object. See change log for 
      details.

Other:
    - Made the timer object a little more robust