A simple, fast, and efficient object detector without FPN.

• This repo provides a neat implementation for YOLOF based on Detectron2. A cvpods version can be found in https://github.com/megvii-model/YOLOF.

You Only Look One-level Feature,
Qiang Chen, Yingming Wang, Tong Yang, Xiangyu Zhang, Jian Cheng, Jian Sun

• Our project is developed on detectron2. Please follow the official detectron2 installation.
• Install mish-cuda to speed up the training and inference when using CSPDarkNet-53 as the backbone (optional)

  git clone https://github.com/thomasbrandon/mish-cuda
  cd mish-cuda
  python setup.py build install
  cd ..

• Install YOLOF by:

  python setup.py develop

• Then link your dataset path to datasets

  cd datasets/
  ln -s /path/to/coco coco

• Download the pretrained model in OneDrive or in the Baidu Cloud with code qr6o to train with the CSPDarkNet-53 backbone (optional)

  mkdir pretrained_models
  # download the `cspdarknet53.pth` to the `pretrained_models` directory

• Train with yolof

  python ./tools/train_net.py --num-gpus 8 --config-file ./configs/yolof_R_50_C5_1x.yaml

• Test with yolof

  python ./tools/train_net.py --num-gpus 8 --config-file ./configs/yolof_R_50_C5_1x.yaml --eval-only MODEL.WEIGHTS /path/to/checkpoint_file

• Note that there might be API changes in future detectron2 releases that make the code incompatible.

The models listed below can be found in this onedrive link or in the BaiduCloud link with code qr6o. The FPS is tested on a 2080Ti GPU. More models will be available in the near future.

• Note that, the speed reported in this repo is 2~3 FPS faster than the one reported in the cvpods version.

If you find this project useful for your research, please use the following BibTeX entry.

@inproceedings{chen2021you,
  title={You Only Look One-level Feature},
  author={Chen, Qiang and Wang, Yingming and Yang, Tong and Zhang, Xiangyu and Cheng, Jian and Sun, Jian},
  booktitle={IEEE Conference on Computer Vision and Pattern Recognition},
  year={2021}
}