I apologize for the extraneous code, as I did not end up having time to properly cleanup my codebase. Too many courses, not enough sleep :,(

I do my best here to specify the relevant code for this course project but because of the many turns this project took during its development, there is some extra code which ended up not being relevant to my results. I will create a cleaned up branch without any modifications whatsoever to the final project code and put a link here, but leave this here (besides the link addition) as it was at the time of submission on April 21st for transparency.

EDIT (Apr 23): Please see the course-proj-cleanup branch of this repo for a much improved README, much clearer file names, and the removal of some outdated files. No actual code was modified in doing so, and of course the file histories can be checked to verify this. While I urge my reviewers to at least read that README instead of this one (if they dont use that branch to evaluate), I've left this one here as is for transparency. Sorry for the mess.

The original videos cannot be shared due to privacy concerns, but I have uploaded the extracted keypoints:

• 'data/body/2d_proposals/tug_stand_walk_sit_CH006_CH007_2D_kpts.pickle' contains the filtered 2D backbone pose predictions (MVP-3D lifter training dataset)
• 'data/body/3d_time_series/CH006_tug_stand_walk_sit.pickle' contains the MVP-3D predictions on the TUG data

• create_PD_2d_dataset.py uses AlphaPose to get the 2D backbone poses on the UPDRS video data and/or filters the 2D poses to create the MVP-3D lifter training dataset
  • 'data/body/body_dataset.py' contains the code for filtering the 2D poses (filter_alphapose_results() and filter_ap_detections())
  • 'utils/alphapose_filtering.py' contains the pixel-space filter definitions
• train_PD_bodylifter.py trains MVP-3D on the desired UPDRS task (Oval or TUG) and also produces the MVP-3D TUG data predictions

• 'uncertnet_experiments.py' contains high level control for training and evaluating UncertNet
• 'uncertnet/uncertnet.py' defines the UncerNet and its training
• 'uncertnet/dataset.py' helps load and handle data for the UncertNet
• 'data/body/h36m/uncertnet' contains the numpy data files for training and testing

3D body extraction from videos of subjects performing UPDRS tasks & post-processing to evaluate their motion features and give UPDRS scores (in the future).

2D Proposal Network: AlphaPose (Halpe Dataset, 26 keypoints)

1. Clone the AlphaPose repo into "auto_UPDRS":

cd auto_UPDRS
git clone https://github.com/MVIG-SJTU/AlphaPose.git

1. Download the AlphaPose pretrained model weights halpe26_fast_res50_256x192.pth (see AlphaPose repo instructions) and place into your "auto_UPDRS/AlphaPose/pretrained_models" directory:

2. Verify that AlphaPose sample "demo_inference.py" script works on their example data:

cd auto_UPDRS/AlphaPose
python3 scripts/demo_inference.py --cfg configs/halpe_26/resnet/256x192_res50_lr1e-3_1x.yaml --checkpoint pretrained_models/halpe26_fast_res50_256x192.pth --indir examples/demo/ --save_img

3. Get the pretrained 3D lifter models from PD_Gait_labeling. Currently we only use "model_lifter.pt", the 3D pose model fine-tuned on our PD walking data using 2 view cameras:

cd auto_UPDRS/model_checkpoints/Mohsens
pip install --upgrade --no-cache-dir gdown
gdown --folder --id 1GVjvla21_oXL4KpSbsEAPXRS7N93W1d5?usp=share_link