itijyou / gaitrecog

PyTorch re-implementation of 'A comprehensive study on cross-view gait based human identification with deep CNNs'

Gait Recognition

PyTorch re-implementation of some models in the paper A comprehensive study on cross-view gait based human identification with deep CNNs.

Usage

0. Download our prepared Casia-B data.

1. Install the fastai library.

2. Clone this repo.

3. Extract the data and arrange the directories as follows.

   somedir
   |-- data
   |   `-- casiab-nm
   `-- gaitrecog
       |-- train_on_gei.py
       `-- ...
   

4. Go to somedir/gaitrecog, and either use the jupyter notebook train_on_gei.ipynb or run the following from a terminal.

   python train_on_gei.py --model lb --task tr --split tv
   python train_on_gei.py --model lb --task ts --split ts --trained casiab-nm_lb_sgd-0.01-0.9-0.0005_st-14-96_bs128_tv_239

Note

Sometimes the training never converges due to bad initialization. In this case, stop the training and restart it.