Unsupervised Pre-training for Person Re-identification (LUPerson).
The repository is for our CVPR2021 paper Unsupervised Pre-training for Person Re-identification.
LUPerson is currently the largest unlabeled dataset for Person Re-identification, which is used for Unsupervised Pre-training. LUPerson consists of 4M images of over 200K identities and covers a much diverse range of capturing environments.
LUPerson can only be used for research, commercial usage is forbidden.
Details can be found at ./LUP.
| Model | path |
|---|---|
| ResNet50 | R50 |
| ResNet101 | R101 |
| ResNet152 | R152 |
For MGN with ResNet50:
| Dataset | mAP | cmc1 | path |
|---|---|---|---|
| MSMT17 | 66.06/79.93 | 85.08/87.63 | MSMT |
| DukeMTMC | 82.27/91.70 | 90.35/92.82 | Duke |
| Market1501 | 91.12/96.16 | 96.26/97.12 | Market |
| CUHK03-L | 74.54/85.84 | 74.64/82.86 | CUHK03 |
These numbers are a little different from those reported in our paper, and most are slightly better.
For MGN with ResNet101:
| Dataset | mAP | cmc1 | path |
|---|---|---|---|
| MSMT17 | 68.41/81.12 | 86.28/88.27 | - |
| DukeMTMC | 84.15/92.77 | 91.88/93.99 | - |
| Market1501 | 91.86/96.21 | 96.56/97.03 | - |
| CUHK03-L | 75.98/86.73 | 75.86/84.07 | - |
The numbers are in the format of without RR/with RR.
If you find this code useful for your research, please cite our paper.
@article{fu2020unsupervised,
title={Unsupervised Pre-training for Person Re-identification},
author={Fu, Dengpan and Chen, Dongdong and Bao, Jianmin and Yang, Hao and Yuan, Lu and Zhang, Lei and Li, Houqiang and Chen, Dong},
journal={Proceedings of the IEEE conference on computer vision and pattern recognition},
year={2021}
}
We extend our LUPerson to LUPerson-NL with Noisy Labels which are generated from tracking algorithm, Please check for our CVPR22 paper Large-Scale Pre-training for Person Re-identification with Noisy Labels. And LUPerson-NL dataset is available at https://github.com/DengpanFu/LUPerson-NL
LUPerson and LUPerson-NL are used by some work and have obtained very good performance.