Awesome

PAST

Self-Training With Progressive Augmentation for Unsupervised Cross-Domain Person Re-identification (Accepted by ICCV19)

[PDF-CameraReady] [Project]

This code is ONLY released for academic use.

This version is ONLY containing testing process. The training code will be released later.

Pipeline

Installation

Python 3.6.5
Pytorch 0.4.1
Torchvision 0.2.1
Please refer to requirements.txt for the other packages with the corresponding versions.

Preparation

Run git clone https://github.com/zhangxinyu-xyz/PAST-ReID.git
Prepare dataset

a. Download datasets: Market-1501, DukeMTMC-reID, CUHK03

b. Move them to $PAST/data/

c. Insure the data folder like the following structure (otherwise you should modify the data path in $PAST/reid/datasets/[DATANAME].py):

$PAST/data
    Market-1501-v15.09.15
        bounding_box_train
        bounding_box_test
        query
    DukeMTMC-reID
        bounding_box_train
        bounding_box_test
        query
    cuhk03-np
        detected
            bounding_box_train
            bounding_box_test
            query

Prepare initial files

a. Download initial model: source-M.pth.tar (trained on Market-1501), source-D.pth.tar (trained on DukeMTMC-reID). Move them to $PAST/initialization/pretrained_model/

b. Download initial feature: D-M_M-t-feature.mat (directly transfer from DukeMTMC-reID to Market-1501), M-D_D-t-feature.mat (directly transfer from Market-1501 to DukeMTMC-reID). Move them to $PAST/initialization/initial_feature/

c. Download initial distmat: D-M_M-t-rerank-distmat.mat (directly transfer from DukeMTMC-reID to Market-1501, after re-rank), M-D_D-t-rerank-distmat.mat (directly transfer from Market-1501 to DukeMTMC-reID, after re-rank). Move them to $PAST/initialization/initial_distmat/

d. If you just want to test our method, you can download our model: D-M_best-model.pth.tar (transfer from DukeMTMC-reID to Market-1501), M-D_best-model.pth.tar (transfer from Market-1501 to DukeMTMC-reID). Move them to $PAST/best_model/

Train

Will be released later. Please wait.

Test

You can simply run test_*.sh file for the transferring testing process.

sh test_D2M.sh  ### from Duke to Market1501
sh test_M2D.sh  ### from Market1501 to Duke

We shall get about 78.38%/54.62% rank-1/mAP on Market-1501 (from DukeMTMC-reID to Market-1501) and 72.35%/54.26% rank-1/mAP on DukeMTMC-reID (from Market-1501 to DukeMTMC-reID).

Results (rank1/mAP)

Model	M-->D	D-->M	C-->M	C-->D
PCB (Direct Transfer)	42.73(25.70)	57.57(29.01)	51.43(27.28)	29.40(16.72)
PCB-R (+Re-rank)	49.69(39.38)	59.74(41.93)	55.91(38.95)	35.19(26.89)
PCB-R-CTL+RTL (+Conservative Stage)	71.63(52.05)	74.26(50.59)	77.70(54.36)	65.71(46.58)
PCB-R-PAST (+Promoting Stage)	72.35(54.26)	78.38(54.62)	79.48(57.34)	69.88(51.79)

References

[1] Our code is conducted based on PCB_RPP_for_reID.

[2] Beyond Part Models: Person Retrievalwith Refined Part Pooling(and A Strong Convolutional Baseline), ECCV2018

[3] Self-training With progressive augmentation for unsupervised cross-domain person re-identification, ICCV2019

Citation

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

@inproceedings{zhang2018selftraining,
title={Self-training With progressive augmentation for unsupervised cross-domain person re-identification},
author={Zhang, Xinyu and Cao, Jiewei and Shen, Chunhua and You, Mingyu},
booktitle ={ICCV},
year={2019}
}

Contact

If you have any questions, please do not hesitate to contact us.

Chunhua Shen

Xinyu Zhang