Pytorch-Person-REID-Baseline-Bag-of-Tricks: A Python repository from upgirlnana

Pytorch-Person-ReID-Baseline-Bag-of-Tricks

Introduction: This project re-implements the strong person re-identification baseline: Bag of Tricks (paper and official code).
Advantage: This project provides a more simple and clear implementation by only using the best parameters and removing lots of unnecessary modules.
Acknowledge: This project refers the official code, if you find this project useful, please cite the offical paper.

News

2019-06-18: we add visualization code to show ranked images
2019-05-01: We re-implement PCB and achieve better performance than the offical one. Our implementation can be found here.

Dependencies

Anaconda (Python 2.7)
PyTorch 0.4.0
GPU Memory >= 10G
Memory >= 20G

Dataset Preparation

Market-1501 Dataset and DukeMTMC-reID Dataset
Download and extract both anywhere

Run

# train
python main.py --market_path market_path --duke_path duke_path --output_path output_path/ --mode train

# test, the output_path should be same with that in training process
python main.py --market_path market_path --duke_path duke_path --output_path output_path/ --mode test --resume_test_epoch resume_test_epoch

# visualize the ranked images, the output_path should be same with that in training process
python main.py --market_path market_path --duke_path duke_path --output_path output_path/ --mode visualize --resume_visualize_epoch resume_visualize_epoch

Experiments

1. Tricks we used

2. Settings

We conduct our experiments on 1 GTX1080ti GPU

3. Results (with REA)

Repeat	market2market	market2duke	duke2duke	duke2market
1	0.939 (0.858)	0.290 (0.159)	0.874 (0.767)	0.486 (0.210)
2	0.944 (0.858)	0.295 (0.156)	0.868 (0.765)	0.492 (0.223)
3	0.942 (0.859)	0.281 (0.152)	0.863 (0.765)	0.485 (0.221)
Average	0.942 (0.858)	0.289 (0.156)	0.868 (0.766)	0.488 (0.218)
Paper	0.941 (0.857)	-	0.864 (0.764)