The proposed anchor loss is simple and highly effective with additional serveral epochs fine-tuning after traditional training stage and brings sigfinicant performance boost. It achieves state-of-the-art based on a simple and strong baseline, Bag of Tricks and a Strong Baseline for Deep Person Re-Identification. Check out our technical report for more details.
If you find the technical report or repository is useful, please kindly cite:
@article{chen2020cluster-reid,
title={Cluster-level Feature Alignment for Person Re-identification},
author={Chen, Qiuyu and Zhang, Wei and Fan, Jianping},
journal={arXiv preprint arXiv:2008.06810},
year={2020}
}
- Market1501
- DukeMTMC
- CUHK03
- Tensorboard visiualization
- Fast inference and validation after every training epoch
pip install -r requirements.txt
cd utils/rank_cylib
make all
cd ../../
Download the pretrained models and place under the root directory of this project.
This stage is identical to original training.
An auxilary anchor loss is added to further fine-tune the model trained after stage 1.
| Train Script | Rank@1 | mAP | Pretrained Model |
|---|---|---|---|
| scripts/trainval_market_bnneck.sh | 95.36% | 87.91% | pretrained_models/strong-baseline-market-bnneck-stage2/e159t30071.pth.tar |
| scripts/trainval_market_bnneck_ibn_a.sh | 95.67% | 89.53% | pretrained_models/strong-baseline-duke-bnneck-ibn-a-stage2/e139t32279.pth.tar |
| Train Script | Rank@1 | mAP | Pretrained Model |
|---|---|---|---|
| scripts/trainval_duke_bnneck.sh | 89.14% | 79.48% | pretrained_models/strong-baseline-duke-bnneck-stage2/e159t37434.pth.tar |
| scripts/trainval_duke_bnneck_ibn_a.sh | 91.11% | 81.84% | pretrained_models/strong-baseline-duke-bnneck-ibn-a-stage2/e139t32279.pth.tar |
Some parts of this repo is taking code repositories beblow as references:
Hugh thanks to the code maintainers of the above repositories.