MoCha Baseline Bundle

This repository is a submit-ready baseline for MoCha2026 - Benchmark and Challenge on Parkinsonian Gait @ ECCV2026.

The challenge evaluates methods that predict MDS-UPDRS gait severity from canonicalized SMPL motion sequences. The CodaBench evaluator imports the submission entry point, calls predict(data), and handles writing prediction files for scoring.

Repository Layout

run.py                    # Required CodaBench entry point
metadata.yaml             # No-op command required by CodaBench
submission/               # Baseline model and preprocessing wrapper
model/                    # Motion encoder architecture
data/preprocessing/       # SMPL-to-motion-feature conversion utilities
human_body_prior/         # Minimal SMPL body model utilities used by preprocessing
chumpy/                   # Compatibility shim for legacy SMPL pickles
weights/                  # Baseline checkpoints and preprocessing statistics

Only run.py is part of the challenge API. The other folders are one example implementation and can be replaced by your own model code.

Expected Input Format

CodaBench calls:

predict(data)

where the unseen challenge data follows this format:

data[subject_id][walk_id] = {
    "pose": np.ndarray,   # shape (T, 72), SMPL axis-angle pose
    "trans": np.ndarray,  # shape (T, 3), global translation
    "beta": np.ndarray,   # shape (1, 10), all zeros for privacy
    "fps": int,           # original collection frame rate
}

The baseline handles FPS conversion internally before inference. If you replace the model, make sure your preprocessing converts each sample to the frame rate expected by your method.

Required Output Format

Your predict function must return a nested Python dict:

predictions[subject_id][walk_id] = predicted_class

where predictions is a dictionary of dictionaries and predicted_class is an integer in {0, 1, 2, 3}.

You do not need to write predictions.json inside CodaBench. The official ingestion program serializes the dictionary returned by predict(data).

Preparing a CodaBench Submission

Create a zip whose root contains run.py, metadata.yaml, and any files or folders needed by your method. The zip should contain the contents of your submission folder, not the submission folder itself.

From inside your submission folder, run:

cd /path/to/your_submission_folder
zip -r ../your_submission.zip . \
  -x "*/__pycache__/*" "*.DS_Store" ".git/*"

Before uploading, you can check the zip root:

unzip -l ../your_submission.zip | head

You should see run.py and metadata.yaml at the top level of the archive, not nested under a project folder.

Upload the resulting zip to the MoCha2026 CodaBench challenge.

Adapting This Baseline

• Keep run.py as a small CodaBench wrapper.
• Replace submission/baseline_model.py with your own model loading and prediction logic.
• Replace submission/preprocess.py if your method uses a different motion representation.
• Include all model weights and pure-Python helper code needed at inference time.
• Do not rely on downloading packages, weights, or data during evaluation.

Model Files

The baseline weights are stored with Git LFS because one checkpoint is larger than GitHub's regular file-size limit. After cloning, run:

git lfs pull

if the large weight files were not downloaded automatically.

Training Data

Participants can train on the CARE-PD dataset:

• Website: https://neurips2025.care-pd.ca/
• Hugging Face: https://huggingface.co/datasets/vida-adl/CARE-PD
• Canonicalized SMPL pickles: https://huggingface.co/datasets/vida-adl/CARE-PD/tree/main/Canonicalized_SMPL_pickles

Contact

For challenge questions, use the CodaBench forum or contact the organizers:

mocha.eccv@gmail.com