TPH-YOLOv5

This repo is the implementation of "TPH-YOLOv5: Improved YOLOv5 Based on Transformer Prediction Head for Object Detection on Drone-Captured Scenarios".
On VisDrone Challenge 2021, TPH-YOLOv5 wins 4th place and achieves well-matched results with 1st place model.
You can get VisDrone-DET2021: The Vision Meets Drone Object Detection Challenge Results for more information.

Install

$ git clone https://github.com/cv516Buaa/tph-yolov5
$ cd tph-yolov5
$ pip install -r requirements.txt

Convert labels

VisDrone2YOLO_lable.py transfer VisDrone annotiations to yolo labels.
You should set the path of VisDrone dataset in VisDrone2YOLO_lable.py first.

$ python VisDrone2YOLO_lable.py

Inference

• Datasets : VisDrone
• Weights (PyTorch v1.10):
  • yolov5l-xs-1.pt: | Baidu Drive(pw: vibe). | Google Drive |
  • yolov5l-xs-2.pt: | Baidu Drive(pw: vffz). | Google Drive |

val.py runs inference on VisDrone2019-DET-val, using weights trained with TPH-YOLOv5.
(We provide two weights trained by two different models based on YOLOv5l.)

$ python val.py --weights ./weights/yolov5l-xs-1.pt --img 1996 --data ./data/VisDrone.yaml
                                    yolov5l-xs-2.pt
--augment --save-txt  --save-conf --task val --batch-size 8 --verbose --name v5l-xs

Ensemble

If you inference dataset with different models, then you can ensemble the result by weighted boxes fusion using wbf.py.
You should set img path and txt path in wbf.py.

$ python wbf.py

Train

train.py allows you to train new model from strach.

$ python train.py --img 1536 --adam --batch 4 --epochs 80 --data ./data/VisDrone.yaml --weights yolov5l.pt --hy data/hyps/hyp.VisDrone.yaml --cfg models/yolov5l-xs-tph.yaml --name v5l-xs-tph

Description of TPH-YOLOv5 and citation

• https://arxiv.org/abs/2108.11539
• https://openaccess.thecvf.com/content/ICCV2021W/VisDrone/html/Zhu_TPH-YOLOv5_Improved_YOLOv5_Based_on_Transformer_Prediction_Head_for_Object_ICCVW_2021_paper.html

If you have any question, please discuss with me by sending email to [email protected]
If you find this code useful please cite:

@inproceedings{zhu2021tph,
  title={TPH-YOLOv5: Improved YOLOv5 Based on Transformer Prediction Head for Object Detection on Drone-captured Scenarios},
  author={Zhu, Xingkui and Lyu, Shuchang and Wang, Xu and Zhao, Qi},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={2778--2788},
  year={2021}
}

References

Thanks to their great works

• ultralytics/yolov5
• SwinTransformer
• WBF