Description

The code is for the task of gait recognition (or person identification) on the MVLP gait dataset. The network employs two Inception modules together with 2 additional convolutions and a fully-connected layer.

• The input is a pair of GEIs concatenated along the channel dimension, i.e. the shape of input tensor is b * h * w * 2 with batch size b.
• The output indicates a probability (provided by a sigmoid) that the two input GEIs are come from the same subject.

This work is to support a project of my friend.

Require

• Numpy
• Pytorch
• Torchvision

Usage

python main.py --task training --epoch 30 --scale 0.25 --angle 90 --batch 12

• --task: "training" or "evaluation"
• --batch: batch size
• --epoch: range of epochs for training or a specific epoch for evaluation
• --scale: parameter scaling the input spatial dimensions
• --angle: the angle of interest (the gallery and probe angles must be the same)

Notice

• Path to the dataset can be assigned in the file main.py.
• This code applies for a single angle, i.e. it considers only a camera angle in both training and evaluation stages. Experiments on different angles are independent.
• The training can be resumed by specifying the epoch range. For example, use --epoch 10-20 to load the model at the 10th epoch and continue to train until the 20th epoch.
• If the task is "training", using --epoch 10 means --epoch 0-10.

Example

Training a model with GEIs of size 32 * 22 (original size in MVLP gait dataset is 128 * 88) corresponding to the camera angle of 90 degrees for 30 epochs and batch size 12:

python main.py --task training --epoch 30 --scale 0.25 --angle 90 --batch 12

Evaluating rank-1 recognition accuracy of the model at the 30th epoch:

python main.py --task evaluation --epoch 30 --scale 0.25 --angle 90 --batch 12

Result

Highest rank-1 recognition accuracies I have obtained for GEIs of size 32 * 22:

Angle (degree)	0	30	60	90
Accuracy (%)	78.2	90.7	88.0	92.7