STGAL: Space-Time Guided Association Learning For Unsupervised Person Re-Identification

This is the code for paper: "Space-Time Guided Association Learning For Unsupervised Person Re-Identification." IEEE International Conference on Image Processing, 2020.
[project page] [paper]

Requirement

We use Python 3.5, Pytorch 0.4.1 in this project. To install required modules, run:

pip3 install -r requirements.txt

Setup

First, set up database files for following datasets.

• Market-1501: Download the dataset to <PATH_TO_MARKET> and run:

  bash setup.sh Market <PATH_TO_MARKET>
  

• DukeMTMC-ReID: Download the dataset to <PATH_TO_DUKE> and run:

  bash setup.sh DukeReID <PATH_TO_DUKE>
  

Run

Train & Test for Market-1501

1. Pretrain on intra-camera data with SSTT[1] (our triplet loss variant):

   bash market/stage1.sh <STAGE1_MODEL_DIR>
   

   * <STAGE1_MODEL_DIR>: Dir to save trained models

2. Construct traveling patterns & label estiamtion:

   bash market/stage1_fusion.sh <STAGE1_CKPT_PATH>
   

   * <STAGE1_CKPT_PATH>: A trained model from the previous step

3. Training:

   bash market/stage2.sh <STAGE2_MODEL_DIR> <STAGE1_CKPT_PATH>
   

   * <STAGE2_MODEL_DIR>: Dir to save trained models * <STAGE1_CKPT_PATH>: The same trained model used in the previous step

4. Inference:

   bash market/stage2_fusion.sh <STAGE2_CKPT_PATH>
   

   * <STAGE2_CKPT_PATH>: The same trained model used in the previous step CMC rank-{1,5,10,20} accuracy and mAP of the 2 settings should appear on the screen:

   • Without fusion probability: Inference with visual similarity
   • With fusion probability: Inference with association probability (visual & traveling probability)

Train & Test for DukeMTMC-ReID

1. Pretrain on intra-camera data with SSTT[1] (our triplet loss variant):

   bash dukereid/stage1.sh <STAGE1_MODEL_DIR>
   

   * <STAGE1_MODEL_DIR>: Dir to save trained models

2. Construct traveling patterns & label estiamtion:

   bash dukereid/stage1_fusion.sh <STAGE1_CKPT_PATH>
   

   * <STAGE1_CKPT_PATH>: A trained model from the previous step

3. Training:

   bash dukereid/stage2.sh <STAGE2_MODEL_DIR> <STAGE1_CKPT_PATH>
   

   * <STAGE2_MODEL_DIR>: Dir to save trained models * <STAGE1_CKPT_PATH>: The same trained model used in the previous step

4. Inference:

   bash dukereid/stage2_fusion.sh <STAGE2_CKPT_PATH>
   

   * <STAGE2_CKPT_PATH>: The same trained model used in the previous step CMC rank-{1,5,10,20} accuracy and mAP of the 2 settings should appear on the screen:

   • Without fusion probability: Inference with visual similarity
   • With fusion probability: Inference with association probability (visual & traveling probability)

Reference

1. Minxian Li, Xiatian Zhu, and Shaogang Gong, “Unsupervised person re-identification by deep learning tracklet association,” in ECCV, 2018, pp. 737–753.

Citation

@inproceedings{wu2020stgal,
	title={Space-Time Guided Association Learning For Unsupervised Person Re-Identification},
	author={Wu, Chih-Wei and Liu, Chih-Ting and Tu, Wei-Chih and Tsao, Yu and Wang, Yu-Chiang Frank and Chien, Shao-Yi},
	booktitle={2020 IEEE International Conference on Image Processing (ICIP)},
	pages={2261--2265},
	year={2020},
	organization={IEEE}
}