Challenge Instructions

To participate in the MoCha challenge:

• Review the CARE-PD data and benchmark resources.
• Train your model using the released CARE-PD training data.
• Package your model weights and code when submissions open.
• Submit your package to the MoCha competition website.
• We run the evaluation and update the leaderboard.

The challenge is based on CARE-PD, a multi-site anonymized clinical dataset for Parkinson's disease gait assessment. CARE-PD provides harmonized 3D SMPL gait meshes collected across multiple cohorts and clinical centers. Each sequence represents a walking trial as privacy-preserving body motion rather than raw video, with clinical gait labels available for model development.

The task is to build a model that predicts the Parkinsonian gait severity score (UPDRS-III) from these motion sequences. The goal is not only to do well on CARE-PD multi-site data, but also to generalize to new clinical sites that were not seen during training.

For more information on dataset details, please visit the CARE-PD website and read the CARE-PD benchmark paper. The public CARE-PD dataset card is available on Hugging Face and includes details on the dataset format. For helper code, you can also take a look at the CARE-PD benchmark GitHub page.

For the MoCha challenge, submissions will be evaluated on harmonized test sequences from additional unseen sites. The primary leaderboard metric is Macro-F1.

Submission is intended to be straightforward: teams will submit their model weights and code in a specific format to the MoCha competition website as a zip or container file. We will evaluate the submitted model on our unseen test sets and update the leaderboard.

The exact packaging format, submission template, and competition website link will be announced soon.

For any enquiries, please email us at mocha.eccv@gmail.com.