GaitGraph

Check out our new version GaitGraph2!

With support for OUMVLP-Pose Dataset, In-Memory Datasets & PyTorch Lightning

Towards a Deeper Understanding of Skeleton-based Gait Recognition

This repository contains the PyTorch code for:

GaitGraph: Graph Convolutional Network for Skeleton-Based Gait Recognition

Torben Teepe, Ali Khan, Johannes Gilg, Fabian Herzog, Stefan Hörmann

Quick Start

Prerequisites

• Python >= 3.6
• CUDA >= 10

First, create a virtual environment or install dependencies directly with:

pip3 install -r requirements.txt

Data preparation

The extraction of the pose data from CASIA-B can either run the commands bellow or download the preprocessed data using:

cd data
sh ./download_data.sh

Optional: If you choose to run the preprocessing, download the dataset and run the following commands.

# Download required weights
cd models
sh ./download_weights.sh

# Copy extraction script
# <PATH_TO_CASIA-B> should be something like: /home/ ... /datasets/CASIA_Gait_Dataset/DatasetB
cd ../data
cp extract_frames.sh <PATH_TO_CASIA-B>

cd <PATH_TO_CASIA-B>
mkdir frames
sh extract_frames.sh
cd frames
find . -type f -regex ".*\.jpg" -print | sort | grep -v bkgrd > ../casia-b_all_frames.csv
cp ../casia-b_all_frames.csv <PATH_TO_REPO>/data

cd <PATH_TO_REPO>/src
export PYTHONPATH=${PWD}:$PYTHONPATH

cd preparation
python3 prepare_detection.py <PATH_TO_CASIA-B> ../../data/casia-b_all_frames.csv ../../data/casia-b_detections.csv
python3 prepare_pose_estimation.py  <PATH_TO_CASIA-B> ../../data/casia-b_detections.csv ../../data/casia-b_pose_coco.csv
python3 split_casia-b.py ../../data/casia-b_pose_coco.csv --output_dir ../../data

Train

To train the model you can run the train.py script. To see all options run:

cd src
export PYTHONPATH=${PWD}:$PYTHONPATH

python3 train.py --help

Check experiments/1_train_*.sh to see the configurations used in the paper.

Optionally start the tensorboard with:

tensorboard --logdir=save/casia-b_tensorboard 

Evaluation

Evaluate the models using evaluate.py script. To see all options run:

python3 evaluate.py --help

Main Results

Top-1 Accuracy per probe angle excluding identical-view cases for the provided models on CASIA-B dataset.

The pre-trained model is available here.

Licence & Acknowledgement

GaitPose itself is released under the MIT License (see LICENSE).

The following parts of the code are borrowed from other projects. Thanks for their wonderful work!

• Object Detector: eriklindernoren/PyTorch-YOLOv3
• Pose Estimator: HRNet/HRNet-Human-Pose-Estimation
• ST-GCN Model: yysijie/st-gcn
• ResGCNv1 Model: yfsong0709/ResGCNv1
• SupCon Loss: HobbitLong/SupContrast

Citing GaitGraph

If you use GaitGraph, please use the following BibTeX entry.

@inproceedings{teepe2021gaitgraph,
  author={Teepe, Torben and Khan, Ali and Gilg, Johannes and Herzog, Fabian and H\"ormann, Stefan and Rigoll, Gerhard},
  booktitle={2021 IEEE International Conference on Image Processing (ICIP)}, 
  title={Gait{G}raph: Graph Convolutional Network for Skeleton-Based Gait Recognition}, 
  year={2021},
  pages={2314-2318},
  doi={10.1109/ICIP42928.2021.9506717}
}