Gait Recognition in the Wild: A Benchmark

GREW-BENCHMARCH

This repository contains the code for our ICCV 2021 paper Gait Recognition in the Wild: A Benchmark

GREW-BENCHMARK project is mainly maintained By Xianda Guo and Zheng Zhu.

For all main contributors, please check contributing.

Dataset

Download GREW datasets.

You can decompress the compressed package with the following command

unzip -P password train.zip (password 是要解压的密码 )
tar -xzvf train.tgz
cd train
ls *.tgz | xargs -n1 tar xzvf

Folder structure after decompression is as follows:

├── make_pose(Silhouettes, Gait Energy Images (GEIs) , 2D and 3D poses)
    ├── train
        ├── 00001
            ├── 4XPn5Z28
                ├── 00001.png
                ├── 00001_2d_pose.txt
                ├── 00001_3d_pose.txt
            ├── 4XPn5Z28_gei.png
    ├── test
        ├── gallery
            ├── 00001
                ├── 79XJefi8
                    ├── 00001.png
                    ├── 00001_2d_pose.txt
                    ├── 00001_3d_pose.txt
                ├── 79XJefi8_gei.png
        ├── probe
            ├── 01DdvEHX
                ├── 00001.png
                ├── 00001_2d_pose.txt
                ├── 00001_3d_pose.txt
            ├── 01DdvEHX_gei.png
    ├── distractors
        ├── edXJefi8
            ├── 00001.pn
            ├── 00001_2d_pose.txt
            ├── 00001_3d_pose.txt
        ├── edXJefi8_gei.png
        
├── flow(optical flow)
    ├── train
        ├── 00001
            ├── 4XPn5Z28
                ├── 00001_00002_flow.png
    ├── test
        ├── gallery
            ├── 00001
                ├── 79XJefi8
                    ├── 00001_00002_flow.png
        ├── probe
            ├── 01DdvEHX
                ├── 00001_00002_flow.png
    ├── distractors
        ├── edXJefi8
            ├── 00001_00002_flow.png

!!! ATTENTION !!! ATTENTION !!! ATTENTION !!!

The Silhouettes and Gait Energy Images (GEIs) has been resized to 64*64. And 2D and 3D poses are obtained from the original image without changing the shape.

The header of *_2d_pose.txt is as follows:

orl_image_w,orl_image_h,nose_x,nose_y,nose_conf,left_eye_x,left_eye_y,left_eye_conf,right_eye_x,right_eye_y,right_eye_conf,left_ear_x,left_ear_y,left_ear_conf,right_ear_x,right_ear_y,right_ear_conf,left_shoulder_x,left_shoulder_y,left_shoulder_conf,right_shoulder_x,right_shoulder_y,right_shoulder_conf,left_elbow_x,left_elbow_y,left_elbow_conf,right_elbow_x,right_elbow_y,right_elbow_conf,left_wrist_x,left_wrist_y,left_wrist_conf,right_wrist_x,right_wrist_y,right_wrist_conf,left_hip_x,left_hip_y,left_hip_conf,right_hip_x,right_hip_y,right_hip_conf,left_knee_x,left_knee_y,left_knee_conf,right_knee_x,right_knee_y,right_knee_conf,left_ankle_x,left_ankle_y,left_ankle_conf,right_ankle_x,right_ankle_y,right_ankle_conf

The header of *_3d_pose.txt is as follows:

orl_image_w,orl_image_h,head_x,head_y,head_z,chin_x,chin_y,chin_z,right_shoulder_x,right_shoulder_y,right_shoulder_z,right_elbow_x,right_elbow_y,right_elbow_z,right_wrist_x,right_wrist_y,right_wrist_z,left_shoulder_x,left_shoulder_y,left_shoulder_z,left_elbow_x,left_elbow_y,left_elbow_z,left_wrist_x,left_wrist_y,left_wrist_z,right_hip_x,right_hip_y,right_hip_z,right_knee_x,right_knee_y,right_knee_z,right_ankle_x,right_ankle_y,right_ankle_z,left_hip_x,left_hip_y,left_hip_z,left_knee_x,left_knee_y,left_knee_z,left_ankle_x,left_ankle_y,left_ankle_z

Requirements

Environment

Python 3.6.*
CUDA 11.1
PyTorch
TorchVision

Install

Create a virtual environment and activate it.

conda create -n grew python=3.6.12
conda activate grew

The code has been tested with PyTorch 1.0 and Cuda 11.1.

conda install pytorch=1.0.0 torchvision=0.2.1
conda install matplotlib tqdm
conda install tensorboard tensorboardX
conda install scipy scikit-image opencv

Train

Train a model by

python train.py

--cache if set as TRUE all the training data will be loaded at once before the training start.
This will accelerate the training.
Note that if this arg is set as FALSE, samples will NOT be kept in the memory
even they have been used in the former iterations. #Default: TRUE

Test

python build_submission.py

Participants must package the submission.csv for submission using zip xxx.zip $CSV_PATH and then upload it to codalab.

Acknowledgements

Part of the code is adopted from previous works: GaitSet, We thank the original authors for their awesome repos.

License

The GREW dataset is freely available for free non-commercial use, and may be redistributed under these conditions.
For commercial queries, contact Zheng Zhu.

Citing

If you find this code useful, please consider to cite our work.

@inproceedings{zhu2021gait,
    title={Gait Recognition in the Wild: A Benchmark},
    author={Zheng Zhu, Xianda Guo, Tian Yang, Junjie Huang, 
        Jiankang Deng, Guan Huang, Dalong Du,Jiwen Lu, Jie Zhou},
    booktitle={IEEE International Conference on Computer Vision (ICCV)},
    year={2021}              
}