Cross-View Gait Based Human Idendification with Deep CNNs

A pytorch implementation of the Local-Bottom Network (LB) in the paper:

Wu, Zifeng, et al. "A comprehensive study on cross-view gait based human identification with deep cnns." IEEE transactions on pattern analysis and machine intelligence 39.2 (2017): 209-226.

Dependency

• python3
• pytorch >= 0.4.0
• visdom.
• opencv

Model

• In src/model.py there are two models: LBNet and LBNet_1. LBNet_1 is more close to the model described in the section 4.2.1 of the original paper. You can select either one. The results are close to each other.

Training

• To train the model, put the CASIA-B dataset silhoutte data under repository
• mkdir snapshot to build the directory for saving models
• goto the src dir and run

  python3 train.py

The model will be saved into the execution dir every 10000 iterations. You can change the interval in train.py.

Monitor the performance

• Install visdom.
• Start the visdom server with python3 -m visdom.server -port 5274 or any port you like (change the port in train.py and test.py)
• Open this URL in your browser: http://localhost:5274 You will see the training loss curve and the validation accuracy curve.

Testing

• goto src dir and run python3 test.py. You can select which snapshot to use by modifying the checkpoint = th.load('../snapshot/snapshot_75000.pth') to other snapshots. Be patient since it takes a long time. The computed similarities will be saved into similarity.npy
• run python3 compute_acc_per_angle to compute the accuracy for each prove view and gallery view. The results will be saved into acc_table.csv

You will get a table like this.

LBNet

LBNet_1