|
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.
|
Examples: C++