Awesome
AICity-reID 2020 (track2)
In this repo, we include the 1st Place submission to AICity Challenge 2020 re-id track (Baidu-UTS submission)
We fuse the models trained on Paddlepaddle and Pytorch. To illustrate them, we provide the two training parts seperatively as following.
- We include the Paddlepaddle training code at Here.
- We include the Pytorch training code at Here.
Performance:
AICITY2020 Challange Track2 Leaderboard
TeamName | mAP | Link |
---|---|---|
Baidu-UTS(Ours) | 84.1% | code |
RuiYanAI | 78.1% | code |
DMT | 73.1% | code |
Trained Models
How to extract features? Please refer to [Here] and there is one simplified version at [Here]. Here we provide one model of the final models.
- SE_imbalance_s1_384_p0.5_lr2_mt_d0_b24+v+aug (AICity 2020) can be downloaded at [GoogleDrive].
The state-of-the-art model achieving 83.41% mAP on VeRi-776, which is based on our TMM paper.
- Training on VehicelNet only (80.91): Res50_imbalance_s1_256_p0.5_lr2_mt_d0_b48 (TMM) can be downloaded at [GoogleDrive].
- Finetuning on VeRi (83.41): ft_Res50_imbalance_s1_256_p0.5_lr1_mt_d0.2_b48_w5 (TMM) can be downloaded at [GoogleDrive].
Extracted Features & Camera Prediction & Direction Prediction:
I have updated the feature. You may download from GoogleDrive or OneDrive (expired by July 1 2022)
├── final_features/
│ ├── features/ /* extracted pytorch feature
│ ├── pkl_feas/ /* extracted paddle feature (include direction similarity)
│ ├── real_query_fea_ResNeXt101_32x8d_wsl_416_416_final.pkl
| ...
│ ├── query_fea_Res2Net101_vd_final2.pkl
│ ├── gallery_cam_preds_baidu.txt /* gallery camera prediction
│ ├── query_cam_preds_baidu.txt /* query camera prediction
| ├── submit_cam.mat /* camera feature for camera similarity calculation
Related Repos:
-
[Vehicle re-ID Paper Collection] https://github.com/layumi/Vehicle_reID-Collection
-
[Person re-ID Baseline] https://github.com/layumi/Person_reID_baseline_pytorch
-
[Person/Vehicle Generation] https://github.com/NVlabs/DG-Net
Citation
Please cite this paper if it helps your research:
@inproceedings{zheng2020going,
title={Going beyond real data: A robust visual representation for vehicle re-identification},
author={Zheng, Zhedong and Jiang, Minyue and Wang, Zhigang and Wang, Jian and Bai, Zechen and Zhang, Xuanmeng and Yu, Xin and Tan, Xiao and Yang, Yi and Wen, Shilei and others},
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops},
pages={598--599},
year={2020}
}
@article{zheng2020beyond,
title={VehicleNet: Learning Robust Visual Representation for Vehicle Re-identification},
author={Zheng, Zhedong and Ruan, Tao and Wei, Yunchao and Yang, Yi and Mei, Tao},
journal={IEEE Transactions on Multimedia (TMM)},
doi={10.1109/TMM.2020.3014488},
note={\mbox{doi}:\url{10.1109/TMM.2020.3014488}},
year={2020}
}
The heatmap visualization is based on
@article{zheng2017discriminatively,
title={A discriminatively learned cnn embedding for person reidentification},
author={Zheng, Zhedong and Zheng, Liang and Yang, Yi},
journal={ACM transactions on multimedia computing, communications, and applications (TOMM)},
volume={14},
number={1},
pages={1--20},
year={2017},
publisher={ACM New York, NY, USA}
}