Overview

Cross-covariate gait recognition aims to identify people under large appearance and acquisition changes, including clothing, carrying conditions, body-shape variations, viewpoints, and other practical covariates.

This project brings together the CCGR benchmark, CCGR-Mini, and CCGR-3D for cross-covariate gait recognition.

Component	Description
CCGR	A large-scale cross-covariate gait benchmark with dense view and covariate annotations.
CCGR-Mini	A compact subset of CCGR for faster research iteration while preserving covariate diversity.
CCGR-3D	An extension of CCGR for studying robust cross-covariate gait recognition with 3D data.

News

• 2026: 3D Gallery for Cross-Covariate Gait Recognition was published in Knowledge-Based Systems.
• 2024: Cross-Covariate Gait Recognition: A Benchmark was accepted by AAAI 2024.

Papers

Paper	Venue	Links
3D Gallery for Cross-Covariate Gait Recognition	Knowledge-Based Systems, 2026	DOI
Cross-Covariate Gait Recognition: A Benchmark	AAAI 2024	AAAI / arXiv

Dataset Access

The released data are provided for non-commercial research use. Please follow the dataset license and cite the corresponding papers when using CCGR, CCGR-Mini, or CCGR-3D.

CCGR

CCGR contains 970 subjects, about 1.6 million sequences, 33 views, and 53 covariates. It provides multiple gait modalities, including RGB, silhouette, parsing, and pose.

Data type	Access
Derived data: silhouette, parsing, pose	Baidu Netdisk ngcw / OneDrive
Raw RGB data	Please sign the RGB data usage agreement and send it to chengyulong@csu.edu.cn.

CCGR-Mini

CCGR-Mini is a lightweight subset of CCGR. It keeps 970 subjects, 47,884 sequences, 53 covariates, and 33 views, while retaining one randomly selected view for each covariate of each subject.

Data type	Access
Derived data: silhouette, parsing, pose	Baidu Netdisk ei8e / OneDrive
Raw RGB data	Baidu Netdisk oyqf / OneDrive

CCGR-3D

CCGR-3D extends CCGR for evaluating cross-covariate gait recognition with 3D data. It provides multiple processed modalities, including silhouette, pose, parsing, SMPL, and skeleton. When using CCGR-3D, please cite both the 2026 KBS paper and the original CCGR benchmark paper.

Data type	Access
Processed data: silhouette, pose, parsing, SMPL, skeleton, point cloud	Baidu Netdisk dci3
Raw RGB data	Please sign the RGB data usage agreement and send it to chengyulong@csu.edu.cn.

Benchmark Results

CCGR

Silhouette input (%)

Method	R1hard	R1easy	R5hard	R5easy
GaitSet	25.3	35.3	46.7	58.9
GaitPart	22.6	32.7	42.9	55.5
GaitGL	23.1	35.2	39.9	54.1
GaitBase	31.3	43.8	51.3	64.4
DeepGaitV2	42.5	55.2	63.2	75.2

Parsing input (%)

Method	R1hard	R1easy	R5hard	R5easy
GaitSet	31.6	42.8	54.8	67.0
GaitPart	29.0	40.9	51.5	64.5
GaitGL	28.4	42.1	46.6	61.4
GaitBase	48.1	62.0	67.7	79.6
DeepGaitV2	58.8	71.8	77.0	87.0

CCGR-Mini

Silhouette input (%)

Method	R1	mAP	mINP
GaitSet	13.77	15.39	5.75
GaitPart	8.02	10.12	3.52
GaitGL	17.51	18.12	6.85
GaitBase	26.99	24.89	9.72
DeepGaitV2	39.37	36.01	16.77

Parsing input (%)

Method	R1	mAP	mINP
GaitSet	18.09	19.18	7.38
GaitPart	10.60	12.29	4.25
GaitGL	22.53	22.58	9.06
GaitBase	38.96	35.48	16.08
DeepGaitV2	50.43	46.53	24.43

Citation

If this project is useful for your research, please cite the following papers.

@article{zou2026_3dgallery,
  title   = {3D Gallery for Cross-Covariate Gait Recognition},
  author  = {Zou, Shinan and Long, Chengyu and Guo, Fan and Tang, Jin},
  journal = {Knowledge-Based Systems},
  volume  = {339},
  pages   = {115576},
  year    = {2026},
  doi     = {10.1016/j.knosys.2026.115576}
}

@article{zou2024_ccgr,
  title   = {Cross-Covariate Gait Recognition: A Benchmark},
  author  = {Zou, Shinan and Fan, Chao and Xiong, Jianbo and Shen, Chuanfu and Yu, Shiqi and Tang, Jin},
  journal = {Proceedings of the AAAI Conference on Artificial Intelligence},
  volume  = {38},
  number  = {7},
  pages   = {7855--7863},
  year    = {2024},
  doi     = {10.1609/aaai.v38i7.28621}
}

License

The dataset is released for non-commercial research use under CC BY-NC-ND unless otherwise specified. Raw RGB data requires a signed usage agreement.

Notes

• In the AAAI 2024 version of the CCGR paper, the positions of R5^easy and R5^hard in Table 2 were reversed. The results shown above use the corrected order.
• In Table 3 of the AAAI 2024 version, the batch size of GaitBase and DeepGaitV2 was recorded incorrectly. The corrected results are shown above. These corrections do not affect the conclusions of the paper.

Contact

For questions about the dataset or paper, please contact zoushinan@csu.edu.cn.