YOLOv4

This is PyTorch implementation of YOLOv4 which is based on ultralytics/yolov3.

• [original Darknet implementation of YOLOv4]

• [ultralytics/yolov5 based PyTorch implementation of YOLOv4].

development log

Expand

• 2020-08-29 - support deformable kernel.
• 2020-08-24 - support channel last training/testing.
• 2020-08-16 - design CSPPRN.
• 2020-08-15 - design deeper model. csp-p6-mish
• 2020-08-11 - support HarDNet. hard39-pacsp hard68-pacsp hard85-pacsp
• 2020-08-10 - add DDP training.
• 2020-08-06 - support DCN, DCNv2. yolov4-dcn
• 2020-08-01 - add pytorch hub.
• 2020-07-31 - support ResNet, ResNeXt, CSPResNet, CSPResNeXt. r50-pacsp x50-pacsp cspr50-pacsp cspx50-pacsp
• 2020-07-28 - support SAM. yolov4-pacsp-sam
• 2020-07-24 - update api.
• 2020-07-23 - support CUDA accelerated Mish activation function.
• 2020-07-19 - support and training tiny YOLOv4. yolov4-tiny
• 2020-07-15 - design and training conditional YOLOv4. yolov4-pacsp-conditional
• 2020-07-13 - support MixUp data augmentation.
• 2020-07-03 - design new stem layers.
• 2020-06-16 - support floating16 of GPU inference.
• 2020-06-14 - convert .pt to .weights for darknet fine-tuning.
• 2020-06-13 - update multi-scale training strategy.
• 2020-06-12 - design scaled YOLOv4 follow ultralytics. yolov4-pacsp-s yolov4-pacsp-m yolov4-pacsp-l yolov4-pacsp-x
• 2020-06-07 - design scaling methods for CSP-based models. yolov4-pacsp-25 yolov4-pacsp-75
• 2020-06-03 - update COCO2014 to COCO2017.
• 2020-05-30 - update FPN neck to CSPFPN. yolov4-yocsp yolov4-yocsp-mish
• 2020-05-24 - update neck of YOLOv4 to CSPPAN. yolov4-pacsp yolov4-pacsp-mish
• 2020-05-15 - training YOLOv4 with Mish activation function. yolov4-yospp-mish yolov4-paspp-mish
• 2020-05-08 - design and training YOLOv4 with FPN neck. yolov4-yospp
• 2020-05-01 - training YOLOv4 with Leaky activation function using PyTorch. yolov4-paspp

Pretrained Models & Comparison

Model	Test Size	APval	AP50val	AP75val	APSval	APMval	APLval	cfg	weights
YOLOv4	672	47.7%	66.7%	52.1%	30.5%	52.6%	61.4%	cfg	weights
YOLOv4pacsp-s	672	36.6%	55.5%	39.6%	21.2%	41.1%	47.0%	cfg	weights
YOLOv4pacsp	672	47.2%	66.2%	51.6%	30.4%	52.3%	60.8%	cfg	weights
YOLOv4pacsp-x	672	49.3%	68.1%	53.6%	31.8%	54.5%	63.6%	cfg	weights
YOLOv4pacsp-s-mish	672	38.6%	57.7%	41.8%	22.3%	43.5%	49.3%	cfg	weights
YOLOv4pacsp-mish	672	48.1%	66.9%	52.3%	30.8%	53.4%	61.7%	cfg	weights
YOLOv4pacsp-x-mish	672	50.0%	68.5%	54.4%	32.9%	54.9%	64.0%	cfg	weights

Requirements

pip install -r requirements.txt

※ For running Mish models, please install https://github.com/thomasbrandon/mish-cuda

Training

python train.py --device 0 --batch-size 16 --img 640 640 --data coco.yaml --cfg cfg/yolov4-pacsp.cfg --weights '' --name yolov4-pacsp

Testing

python test.py --img 640 --conf 0.001 --batch 8 --device 0 --data coco.yaml --cfg cfg/yolov4-pacsp.cfg --weights weights/yolov4-pacsp.pt

Citation

@article{bochkovskiy2020yolov4,
  title={{YOLOv4}: Optimal Speed and Accuracy of Object Detection},
  author={Bochkovskiy, Alexey and Wang, Chien-Yao and Liao, Hong-Yuan Mark},
  journal={arXiv preprint arXiv:2004.10934},
  year={2020}
}

@inproceedings{wang2020cspnet,
  title={{CSPNet}: A New Backbone That Can Enhance Learning Capability of {CNN}},
  author={Wang, Chien-Yao and Mark Liao, Hong-Yuan and Wu, Yueh-Hua and Chen, Ping-Yang and Hsieh, Jun-Wei and Yeh, I-Hau},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops},
  pages={390--391},
  year={2020}
}

Acknowledgements

• https://github.com/AlexeyAB/darknet
• https://github.com/ultralytics/yolov3
• https://github.com/ultralytics/yolov5