MMGait: Towards Multi-Modal Gait Recognition

MMGait is a large-scale multi-modal benchmark for single-modal, cross-modal, multi-modal, and omni multi-modal gait recognition. It simultaneously captures 13 modalities from multiple sensors, enabling comprehensive cross-modal and multimodal gait analysis.

[CVPR 2026] "MMGait: Towards Multi-Modal Gait Recognition"

📊 Dataset Statistics

Item	Value
Total size	~46 GB (compressed)
Data format	Python pickle (.pkl)
Modalities per sequence	13
Codebase	BNU-IVC/MMGait

🔧 Modalities

Each walking sequence contains the following 13 modalities:

data_in_use follows this fixed order:

[depth, event, heatmap, ir, ir_sils, lidar_depth, lidar_points, pose2d, pose3d, radar4d, radar_depth, rgb, rgb_sils]

📁 Directory Structure

The loader expects the following directory layout:

MMGait_ROOT/
`-- {subject_id}/
    `-- {seq_type}/
        `-- {view_id}/
            |-- depth.pkl
            |-- event.pkl
            |-- heatmap.pkl
            |-- ir.pkl
            |-- ir_sils.pkl
            |-- lidar_depth.pkl
            |-- lidar_points.pkl
            |-- pose2d.pkl
            |-- pose3d.pkl
            |-- radar4d.pkl
            |-- radar_depth.pkl
            |-- rgb.pkl
            `-- rgb_sils.pkl

📥 Download & Usage

The dataset is split into 12 chunks (~4 GB each) under the chunks/ directory. Download and reassemble them as follows:

Python (using huggingface_hub)

from huggingface_hub import list_repo_files, hf_hub_download
import os

repo_id = "TerranceWcy/MMGait"

# 1. Download all chunks (requires approved access)
chunk_files = sorted([f for f in list_repo_files(repo_id, repo_type="dataset") if f.startswith("chunks/")])
local_dir = "./mmgait_chunks"
os.makedirs(local_dir, exist_ok=True)

for f in chunk_files:
    print(f"Downloading {f} ...")
    hf_hub_download(repo_id=repo_id, filename=f, repo_type="dataset", local_dir=local_dir)

# 2. Reassemble into single tar.gz
output_file = "MMGait_Release.tar.gz"
chunk_dir = os.path.join(local_dir, "chunks")
with open(output_file, "wb") as out:
    for f in sorted(os.listdir(chunk_dir)):
        if f.startswith("MMGait_Release.tar.gz.part_"):
            with open(os.path.join(chunk_dir, f), "rb") as chunk:
                out.write(chunk.read())
print(f"Reassembled to {output_file}")

# 3. Extract
import tarfile
with tarfile.open(output_file, "r:gz") as tar:
    tar.extractall()

Command line (using huggingface-cli + cat)

# Download all chunks
huggingface-cli download TerranceWcy/MMGait --repo-type=dataset --local-dir ./mmgait

# Reassemble
cat ./mmgait/chunks/MMGait_Release.tar.gz.part_* > MMGait_Release.tar.gz

# Verify integrity (optional)
gzip -t MMGait_Release.tar.gz

# Extract
tar -xzf MMGait_Release.tar.gz

Note: The 12 chunk files are named MMGait_Release.tar.gz.part_aa through MMGait_Release.tar.gz.part_al. Use cat with sorted order to reassemble correctly.

📜 License

This dataset is released under the Creative Commons Attribution-NonCommercial 4.0 International (CC-BY-NC-4.0) license.

• ✅ You may share and adapt the dataset
• ✅ You must give appropriate credit
• ❌ You may not use the dataset for commercial purposes

This dataset is strictly intended for academic purposes.

🔒 Access Policy (Gated)

This dataset is gated and requires an access application. To request access:

1. Log in to your HuggingFace account
2. Click the "Request access" button on the dataset page
3. Fill in the application form describing your research purpose and institutional affiliation
4. Wait for manual review and approval

Applications are reviewed manually. Please allow several business days for processing.

📄 Citation

If you find this dataset useful in your research, please consider citing our paper:

@article{wang2026mmgait,
  title={MMGait: Towards Multi-Modal Gait Recognition},
  author={Wang, Chenye and Cai, Qingyuan and Hou, Saihui and Li, Aoqi and Huang, Yongzhen},
  journal={arXiv preprint arXiv:2604.15979},
  year={2026}
}

📬 Contact

If you have any questions, please contact chenye.wang@mail.bnu.edu.cn.