Implementation of ICCV2019 paper Beyond Human Parts: Dual Part-Aligned Representations for Person ReID
Codes from this repo can reproduce our results on DukeMTMC-reID.
- Python 3.6
- GPU Memory >= 6G
- Numpy
- Pytorch >= 0.4
- Torchvision >= 0.2.0
Download DukeMTMC-ReID Dataset.
Preparation: You may need our generated human part masks from BaiduCloud. Remember to change the dataset path to your own path in duke.py.
CUHK03 human part masks from BaiduCloud. pwd: q39a
Market-1501 human part masks from BaiduCloud. pwd: uyus
Generated human part masks from Google Drive.
Train a model by
cd scripts
sh resnet50_softmax.shThis model is based on ResNet-50. Input images are resized to 384x128.
Note that results may be better than Table 9 in the paper. (Setting here is batchsize 48 on 1 GPU)
| Method | Rank-1 | Rank-5 | Rank-10 | mAP | Model |
|---|---|---|---|---|---|
| Baseline | 81.10 | 89.59 | 92.19 | 64.87 | BaiduCloud |
| 1 x Latent | 82.92 | 91.03 | 93.49 | 67.09 | BaiduCloud |
| 1 x DPB | 84.83 | 92.28 | 94.08 | 68.62 | BaiduCloud |
@InProceedings{Guo_2019_ICCV,
author = {Guo, Jianyuan and Yuan, Yuhui and Huang, Lang and Zhang, Chao and Yao, Jin-Ge and Han, Kai},
title = {Beyond Human Parts: Dual Part-Aligned Representations for Person Re-Identification},
booktitle = {The IEEE International Conference on Computer Vision (ICCV)},
month = {October},
year = {2019}
}