MMCM: Multimodality-aware Metric using Clustering-based Modes for Probabilistic Human Motion Prediction

This is the official repository for the following paper:

Kyotaro Tokoro, Hiromu Taketsugu, Norimichi Ukita

MMCM: Multimodality-aware Metric using Clustering-based Modes for Probabilistic Human Motion Prediction, WACV 2026

arxiv, suppl

Environment

• Please install the appropriate version of PyTorch (e.g., v.2.3.0, v.2.6.0) for your environment. Then, install the remaining dependencies by running:

pip install -r requirements.txt

Prepare datasets

• Prepare Human3.6M and AMASS following BeLFusion in "./auxiliar".

Prepare weights and so on

• Download weights set, unzip the file, and place it in './compute_mmcm/default_parms'.

Compute MMCM

From numpy output

• You can evaluate predictions saved in NumPy format (.npy).
• Please refer to save_baseline_as_npy.py for how to export your prediction results as an .npy file. By running this script, you can also save the outputs of a very simple baseline predictor.
• The output results for one method (CoMusion) can be downloaded from npy results, and you unzip the file and place it in './baseline_output/comusion/h36m/'. Please note that the resulting zip file is quite large (about 4.5 GB).

# Baseline --> {comusion, belfusion, dlow, and so on}
# Dataset --> {h36m, amass}
python evaluate_baseline.py --pred_path "baseline_output/<Baseline>/<Dataset>/npy/" --data_config_path "compute_mmcm/default_parms/<Dataset>/<Dataset>_config.json" --dataset_name <Dataset>

Base form

• coming soon

Hyperparameter search

If you want to do hyperparameter search on new datases, please use compute_mmcm/parameter_search.py script.

For example,

python compute_mmcm/parameter_search.py --data_config_path compute_mmcm/default_parms/h36m/h36_config.json --stride 25 --frames 103

python compute_mmcm/parameter_search.py --data_config_path compute_mmcm/default_parms/amass/amass_config.json --stride 60 --frames 123