Spatial-Temporal-Pooling-Networks-ReID

Code for our ICCV 2017 paper -- Jointly Attentive Spatial-Temporal Pooling Networks for Video-based Person Re-Identification

If you use this code please cite:

@inproceedings{shuangjiejointly,
  	title={Jointly Attentive Spatial-Temporal Pooling Networks for Video-based Person Re-Identification},
  	author={Shuangjie Xu, Yu Cheng, Kang Gu, Yang Yang, Shiyu Chang and Pan Zhou},
  	booktitle={ICCV},
  	year={2017}
}

Dependencies

The following libaries are necessary:

• torch and its package (nn, nnx, optim, cunn, cutorch, image, rnn , inn). Installation guide
• CUDA support with Nvidia GPU
• Matlab for data preparation

Data Preparation

Download and extract datasets iLIDS-VID, PRID2011 and MARS into the data/ directory. data/iLIDS-VID for example.

Modify and run data/computeOpticalFlow.m with Matlab to generate Optical Flow data. Optical Flow data will be generated in the same dir of your datasets. data/iLIDS-VID-OF-HVP for example.

MARS needs some extra codes to randomly choose two videos for a person (cam1 and cam2). Will release soon.

Evaluation

You can evaluate this network without training. Download weights files for iLIDS-VID or PRID2011 (baidu yun or google drive) into weights/, and run

th reIdEval.lua -dataset 1 -weight /weights/ilids_600_convnet.dat
th reIdEval.lua -dataset 2 -weight /weights/prid_600_convnet.dat

dataset 1 for iLIDS-VID, 2 for PRID2011 and 3 for MARS (release soon)

Training

Run this command for training iLIDS-VID. To train other datasets, change options -dataset.

th reIdTrain.lua -nEpochs 600 -dataset 1 -dropoutFrac 0.6 -sampleSeqLength 16 -samplingEpochs 100 -seed 1 -mode 'spatial_temporal'

The option mode has four type: cnn-rnn, spatial, temporal and spatial_temporal.