3D Local Convolutional Neural Networks for Gait Recognition

Overview

This repository is the official implementation of:

3D Local Convolutional Neural Networks for Gait Recognition

Zhen Huang, Dixiu Xue, Xu Shen, Xinmei Tian, Houqiang Li, Jianqiang Huang, Xian-Sheng Hua

In this work, we present a new building block for 3D CNNs with local information incorporated, termed as 3D local convolutional neural networks. Our local operations can be combined with any existing architectures. We demonstrate the superiority of local operations on the task of gait recognition where 3D local CNN consistently outperforms state-of-the-art models. We hope this work will shed light on more research on introducing simple but effective local operations as submodules of existing convolutional building blocks.

Installation Instructions

• Clone this repo:

git clone git@github.com:aliyun/3D-Local-CNN-for-Gait-Recognition.git
cd 3D-Local-CNN-for-Gait-Recognition

• Create a conda virtual environment and activate it:

conda create -n 3DLocalCNN python=3.6.4 -y
conda activate 3DLocalCNN

• Install CUDA==10.2 with cudnn7 following the official installation instructions
• Install PyTorch==1.7.1 and torchvision==0.8.2 with CUDA==10.2:

conda install pytorch==1.7.1 torchvision==0.8.2 cudatoolkit=10.2 -c pytorch

• Install numpy==1.16.4, yaml, tensorboard, pyyaml, scikit-learn, opencv-python, imageio, matplotlib, seaborn, xarray:

pip3 install numpy==1.16.4, yaml, tensorboard, pyyaml, scikit-learn, opencv-python, imageio, matplotlib, seaborn, xarray

Usage

Data preprocess

Download CASAI raw data to data/CASIA_raw and run python preprocess.py

Demo

(1) To train GaitSet from scratch, run

python main.py --config=configs/GaitSet_CASIA.yaml

(2) To train GaitPart from scratch, run

python main.py --config=configs/GaitPart_CASIA.yaml

(3) To train 3DLocalCNN from scratch, run

python main.py --config=configs/3DLocalCNN_CASIA.yaml

License

• Apache License 2.0