JVTC

Introduction

This is the code of Joint Visual and Temporal Consistency for Unsupervised Domain Adaptive Person Re-Identification ECCV 2020.

Preparation

Requirements: Python=3.6 and Pytorch>=1.0.0

Please refer to ECN to prepare dataset, the file structure is as follow:

JVTC/data    
│
└───Market-1501 OR DukeMTMC-reID
   │   
   └───bounding_box_train
   │   
   └───bounding_box_test
   │   
   └───bounding_box_train_camstyle_merge
   | 
   └───query

"bounding_box_train_camstyle_merge" dir merges the "bounding_box_train" and "bounding_box_train_camstyle" for convenience.

Training and test

We utilize 2 GTX-2080TI GPU for model training.

# Duke to Market-1501 training&evalution
python duke2market_train.py

# Duke to Market-1501 evalution with joint similarity
python duke2market_evaluate_joint_sim.py

# Market-1501 to Duke training&evalution
python market2duke_train.py

# Market-1501 to Duke evalution with joint similarity
python market2duke_evaluate_joint_sim.py

Results

References

• [1] Invariance Matters: Exemplar Memory for Domain Adaptive Person Re-identification. CVPR 2019.

• [2] Spatial-temporal person re-identification. AAAI 2019.

Citation

If you find this code useful in your research, please consider citing:

Contact me

If you have any questions about this code or paper, please contact me at.

Jianing Li