CoMHR

Contrastive Multi-Modal Hypergraph Reasoning for 3D Crowd Mesh Recovery

Official model weights for our ICME 2026 Oral paper.

Paper: https://arxiv.org/abs/2605.13854

Code: https://github.com/SunMH-try/CoMHR

Pretrained Weights

Download CoHMR_weights.zip from this repository.

The pretrained model is also available on Baidu Netdisk (code: k3qd).

Ablation Checkpoints

Additional checkpoints from our ablation experiments are available in the abl_ckpt directory.

Checkpoint	Modalities	Panoptic MPJPE ↓
rgb.pkl	RGB	110.41 mm
depth.pkl	Depth	137.40 mm
pose.pkl	Pose	184.36 mm
multimodal_no_contra.pkl	RGB + Depth + Pose	105.26 mm

The single-modal checkpoints serve as baselines, while multimodal_no_contra.pkl represents the multimodal model with pelvis-depth refinement but without contrastive learning.

The full CoMHR method achieves 104.18 mm MPJPE on Panoptic.

Usage

For model architecture, configuration, dependencies, and inference instructions, please refer to the official GitHub repository.

Citation

If you find our work useful, please consider citing:

@inproceedings{sun2026contrastive,
  title     = {Contrastive Multi-Modal Hypergraph Reasoning for 3D Crowd Mesh Recovery},
  author    = {Sun, Minghao and Xu, Chongyang and Xie, Yitao and Huang, Buzhen and Li, Kun},
  booktitle = {Proceedings of the IEEE International Conference on Multimedia and Expo (ICME)},
  year      = {2026}
}