CARE-PD is a dataset and evaluation benchmark suite for clinical gait analysis in Parkinson’s Disease, released as part of NeurIPS 2025 Datasets & Benchmarks Track submission.
CARE-PD is the largest publicly available archive of 3D mesh gait data for Parkinson's Disease (PD) and the first to include data collected across multiple sites. The dataset aggregates 9 cohorts from 8 clinical sites, including 363 participants spanning a range of disease severity. All recordings—whether from RGB video or motion capture—are unified into anonymized SMPL body gait meshes through a curated harmonization pipeline.
This dataset enables two main benchmarks:
- Supervised clinical score prediction: Estimating UPDRS gait scores from 3D meshes
- Unsupervised motion pretext tasks: Parkinsonian gait representation learning
git clone https://github.com/TaatiTeam/CARE-PD.git
cd CARE-PD
We tested our code on Python 3.9.21 and PyTorch 2.6.0
python -m venv carepd
source carepd/bin/activate
pip install --upgrade pip
pip install -r requirements.txt
pip install torch==2.6.0+cu118 torchvision==0.21.0+cu118 --index-url https://download.pytorch.org/whl/cu118
mkdir -p assets/datasets
Download the CARE-PD datasets from Dataverse and put them in the assets/datasets folder.
You can use smpl_reader to read files and get a summary stats:
python data/smpl_reader.py --dataset PD-GaM BMCLab 3DGait T-SDU-PD DNE E-LC KUL-DT-T T-LTC T-SDU
Download preprocessed h36m formats from Dataverse and put them in the assets/datasets folder.
Rename the folder:
mv assets/datasets/h36m_preprocessed assets/datasets/h36m
You can also preprocess all datasets with the following command but it might take quite some time:
bash scripts/preprocess_smpl2h36m.sh
Download preprocessed HumanML3D formats from Dataverse and put them in the assets/datasets folder.
Rename the folder:
mv assets/datasets/HumanML3D_preprocessed assets/datasets/HumanML3D
You can also preprocess all datasets with the following command but it might take quite some time:
bash scripts/preprocess_smpl2humanml3d.sh
Download preprocessed 6D rotation formats from Dataverse and put them in the assets/datasets folder.
Rename the folder:
mv assets/datasets/6D_preprocessed assets/datasets/6D_SMPL
You can also preprocess all datasets with the following command but it might take quite some time:
bash scripts/preprocess_smpl2sixD.sh
Please also check dataset.md for more information. The dataset directory structure should look:
assets/
└── datasets/
└── 6D_SMPL/
├── 3DGait/
├── ...
├── folds/
├── Other_Datasets/
├── UPDRS_Datasets/
├── h36m/
├── 3DGait/
├── ...
├── HumanML3D/
├── 3DGait/
├── ...
├── 3DGait.pkl
├── BMCLab.pkl
├── DNE.pkl
├── E-LC.pkl
├── KUL-DT-T.pkl
├── PD-GaM.pkl
├── T-LTC.pkl
├── T-SDU-PD.pkl
├── T-SDU.pkl
bash scripts/download_models.sh
Pretrained checkpoints will be downloaded in assets/Pretrained_checkpoints
The output for each experiment will be saved inside a folder in experiment_outs/**experiment_name** specified in the config json file. To see full list for configs please check the config generator for each model inside ./configs folder.
You can run hyperparameter tuning on the BMCLab dataset across all backbone models using:
bash scripts/hypertune_all_models.sh
You can also run a single tuning job manually like this:
python eval_encoder_hypertune.py \
--backbone MODELNAME \
--config CONFIGNAME.json \
--hypertune 1 \
--tune_fresh 1 \
--this_run_num 0 \
--ntrials 50
After hypertuning on BMCLab, you can tune only the number of epochs for each remaining dataset using:
bash scripts/hypertune_epochs_all_datasets.sh
You can also run a single dataset tuning job like:
python run.py \
--backbone MODELNAME \
--config CONFIGNAME.json \
--hypertune 1 \
--tune_fresh 1 \
--ntrials 5 \
--this_run_num 0
You can run final Within-Dataset evaluation on each dataset using:
bash scripts/eval_within_dataset.sh
This script:
- Loads the best hyperparameters from each study
- Retrains the model from scratch on the full training folds
- Evaluates performance in a LOSO setup
- Automatically combines predictions from back and side views (for multi-view models)
- Logs results and confusion matrices to
reports/intra_eval/
You can also run a single dataset evaluation using:
python run.py \
--backbone MODELNAME \
--config CONFIGNAME.json \
--hypertune 0 \
--cross_dataset_test 0 \
--this_run_num 0 \
--num_folds -1
MODELNAME in (potr, momask, motionclip).
python run.py \
--backbone MODELNAME \
--hypertune 0 \
--cross_dataset_test 0 \
--this_run_num 0 \
--num_folds -1 \
--combine_views_preds 1 \
--views_path \
"Hypertune/MODELNAME_CONFIGNAME_backright/0" \
"Hypertune/MODELNAME_CONFIGNAME_sideright/0"
MODELNAME in (motionbert, mixste, poseformerv2, motionagformer).
To run using predefined best configs you can pass directly the best model configs:
python eval_only.py \
--backbone MODELNAME \
--config CONFIGNAME.json \
--hypertune 0 \
--num_folds -1 \
"--tuned_config", "./configs/best_configs_augmented/....json"
After within-dataset testing, you can evaluate how well each model generalizes across datasets.
To run all cross-dataset experiments:
bash scripts/eval_cross_dataset.sh
This script:
- Loads the best hyperparameters from each model's tuning run
- Trains each model on its original dataset
- Tests on all other datasets (automatically handled in code)
- Combines predictions from multiple views for multi-view models
- Logs all outputs to
reports/cross_eval/
To evaluate on a single model and dataset use:
python run.py \
--backbone MODELNAME \
--config CONFIGNAME.json \
--hypertune 0 \
--cross_dataset_test 1 \
--this_run_num 0
MODELNAME in (potr, momask, motionclip).
python run.py \
--backbone MODELNAME \
--hypertune 0 \
--cross_dataset_test 1 \
--combine_views_preds 1 \
--views_path \
"Hypertune/MODELNAME_CONFIGNAME_backright/0" \
"Hypertune/MODELNAME_CONFIGNAME_sideright/0"
MODELNAME in (motionbert, mixste, poseformerv2, motionagformer).
To run using predefined best configs you can pass directly the best model configs:
python eval_only.py \
--backbone MODELNAME \
--config CONFIGNAME.json \
--hypertune 0 \
--cross_dataset_test 1\
"--tuned_config", "./configs/best_configs_augmented/....json"
For Leave-One-Dataset-Out (LODO) evaluation, we first tune the number of training epochs on each dataset excluding the target dataset (i.e., LODO setup).
To run all epoch-tuning jobs for LODO:
bash scripts/hypertune_lodo_epochs.sh
This script:
- Tunes the number of epochs per dataset used in LODO training
- Forces LODO=True using --force_LODO 1
- Uses --exp_name_rigid LODO to name all output folders consistently
- Logs all runs to
reports/hypertune_lodo/
To evaluate on a single model and dataset use:
python run.py \
--backbone MODELNAME \
--config CONFIGNAME.json \
--ntrials 5 \
--this_run_num 0 \
--hypertune 1 \
--tune_fresh 1 \
--force_LODO 1 \
--exp_name_rigid LODO
MODELNAME in (potr, momask, motionclip, motionbert, mixste, poseformerv2, motionagformer).
In this step, we evaluate how well each model generalizes across datasets when trained using a Leave-One-Dataset-Out (LODO) strategy.
Each model is:
- Trained on all datasets except the target
- Evaluated only on the left-out dataset
To run all LODO evaluation jobs:
bash scripts/eval_lodo.sh
All logs are saved in: reports/lodo_eval/
To evaluate on a single model and dataset use:
python run.py \
--backbone MODELNAME \
--config CONFIGNAME.json \
--this_run_num 0 \
--hypertune 0 \
--cross_dataset_test 1 \
--force_LODO 1 \
--exp_name_rigid LODO
MODELNAME in (potr, momask, motionclip).
python run.py \
--backbone MODELNAME \
--hypertune 0 \
--cross_dataset_test 1 \
--force_LODO 1 \
--exp_name_rigid LODO \
--combine_views_preds 1 \
--views_path \
"LODO/MODELNAME_CONFIGNAME_backright_LODO/0" \
"LODO/MODELNAME_CONFIGNAME_sideright_LODO/0"
MODELNAME in (motionbert, mixste, poseformerv2, motionagformer).
To run using predefined best configs you can pass directly the best model configs:
python eval_only.py \
--backbone MODELNAME \
--config CONFIGNAME.json \
--hypertune 0 \
--cross_dataset_test 1\
--force_LODO 1 \
"--tuned_config", "./configs/best_configs_augmented/LODO/....json"
The final evaluation step uses Multi-dataset In-domain Adaptation training under a LOSO setup.
Each model is:
- Trained on all datasets, plus the training portion of the in domain dataset
- Evaluated on the test portion only
- Configured with
--AID 1,--force_LODO 1, and--num_folds -1to reflect this setup
To run all MIDA evaluations:
bash scripts/eval_mida.sh
All logs are stored under: reports/mida_eval
To evaluate on a single model and dataset use:
python run.py \
--backbone MODELNAME \
--config CONFIGNAME.json \
--this_run_num 0 \
--hypertune 0 \
--cross_dataset_test 1 \
--force_LODO 1 \
--AID 1 \
--num_folds -1 \
--exp_name_rigid LODO
MODELNAME in (potr, momask, motionclip).
python run.py \
--backbone MODELNAME \
--hypertune 0 \
--cross_dataset_test 1 \
--force_LODO 1 \
--AID 1 \
--num_folds -1 \
--exp_name_rigid LODO \
--combine_views_preds 1 \
--views_path \
"LODO/MODELNAME_CONFIGNAME_backright_LODO/0" \
"LODO/MODELNAME_CONFIGNAME_sideright_LODO/0"
MODELNAME in (motionbert, mixste, poseformerv2, motionagformer).
We sincerely thank all participating research institutions and subjects who made this dataset possible. We also thank the open-sourcing of these works where some of our code is based on: HumanML3D, joints2smpl, MoMask, MotionBert, MotionBert, MotionCLIP, PoserFormerV2, POTR, MixSTE
If you use CARE-PD in your research, please cite:
[Citation information to be provided upon publication]