Akimoto-Cris / Pytorch_AMOC

AMOC, "Video-based Person Re-identification with Accumulative Motion Context".

Usage

Environment:

• anaconda
• python3.6
• pytorch
• matlab 2019

Dataset

Two options:

• i-LIDS-Vid
• prid

Preprocess

1. modified line 34 in the matlab code computeOpticalFlow.m to your dataset path, then run the matlab code. This produces ground truth for training the MotionNet.
2. create anaconda environment by conda create -n torch python==3.6 pytorch==1.3.0 ignite -c pytorch.
3. activate conda environment and install required libraries: pip install -r requirements.txt.

Training

1. Activate anaconda environment. conda activate torch.

2. (Optional) open a visdom service: python -m visdom.server.

3. Open another terminal/cmd, train the MotionNet sub model:

   python train_motionnet.py -l 1e-4 --saveFileName motionnet -dataset 0.

4. Load the above pretrained motionnet model and train the main model:

   python videoReid.py --train --usePredefinedSplit -l 1e-3 --saveFileName amoc -mp path/to/pretrained_motionnet_model.pth.

Testing

python track_demo.py --source path/to/video_file --tracker KCF/GOTURN/CSRT -p path/to/amoc_weights.pth