New Features, major updates and upgrades:
- new Project Manager GUI (see docs/PROJECT_GUI.md)
- dynamic auto-cropping is available (issue #198): use during analyze_videos: deeplabcut.analyze_videos(path_config_file,[video], dynamic=(True,.1,30)). This dramatically improves the inference speed for typical applications in neuroscience such as open field behaviors where the animal is much smaller than the frame and no prior tracking is necessary.
- Integrated MobileNetV2s (described in Pretraining boosts out-of-domain robustness for pose estimation by Alexander Mathis, Mert Yüksekgönül, Byron Rogers, Matthias Bethge, Mackenzie W. Mathis https://arxiv.org/abs/1909.11229)
- integrated TF based inference code (based on https://arxiv.org/abs/1909.11229).
- Added in new data augmentation options with imgaug see Box 2 in docs (https://github.com/aleju/imgaug) NOTE FOR WINDOWS THIS REQUIRES Shapley https://github.com/AlexEMG/DeepLabCut/issues/446)
- New function, create_training_model_comparison that allows the user to create identical shuffles for multiple network and/or data augmentation testing.
- Updated evaluation metric: one can now evaluate the model at different scales (by passing rescale=True one can evaluate the model at the ‘global_scale’ variable (as set in the test/pose_config.yaml file for a particular project). I.e. every image will be resized according to that scale and prediction will be compared to the resized ground truth. The error will be reported in pixels at rescaled to the original size. I.e. For a [200,200] pixel image evaluated at global_scale=.5, the predictions are calculated on [100,100] pixel images, compared to 1/2*ground truth and this error is then multiplied by 2. The evaluation images are also shown for the original size!
- Retraining now also loads deconvolution weights by default
- dlc scorer name changed, and backwards compatibility maintained. E.g. data for videos like abc.mp4 will be saved as abcDLC_resnet50_xyz.h5 or abcDLC_mobnet35_xyz.h5 (depending on the network)
Bug fixes and minor edits:
- Changed imresize and imread from scipy.misc (deprecated functions); now uses openCV
- Colors changed when manually extracting outliers Extract outlier frames saves images with different color
- Minor edits to ShortenVideo function, now returns path to shortened video
- By default .csv exported when analyzing a video when initializing a human pre-trained network (#427)
- Changes to default values in augmentation with TensorPack; See results here: #426 > Thanks @katierupp! and the from #429 thanks @ppwwyyxx!
- Updated link to human DEMO on COLAB
- updated setup.py dependencies (addresses #434, #433, #441, #440)
- edited and expanded test functions
- expanded docs, updated wiki