AICity-reID 2020 (track2)

In this repo, we include the 1st Place submission to AICity Challenge 2020 re-id track (Baidu-UTS submission)

[Paper] [Video]

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：

• 🚁 Drone-based building re-id [code] [paper]

• [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}
}