AP-loss

The implementation of “Towards accurate one-stage object detection with AP-loss”.

Requirements

Python 2.7
PyTorch 1.3+
Cuda

Installation

Clone this repo

git clone https://github.com/cccorn/AP-loss.git
cd AP-loss

Install the python packages:

pip install pycocotools
pip install opencv-python

Create directories:

mkdir data models results

Prepare Data. You can use

ln -s $YOUR_PATH_TO_coco data/coco
ln -s $YOUR_PATH_TO_VOCdevkit data/voc

The directories should be arranged like:

├── data
│   ├── coco
│   │   ├── annotations
│   │   ├── images
│   │   │   ├── train2017
│   │   │   ├── val2017
│   │   │   ├── test-dev2017
│   ├── voc
│   │   ├── VOC2007
│   │   ├── VOC2012

Prepare the pre-trained models and put them in models like:

├── models
│   ├── resnet50-pytorch.pth
|   ├── resnet101-pytorch.pth

We use the ResNet-50 and ResNet-101 pre-trained models which are converted from here. We also provide the converted pre-trained models at this link.

Training

bash train.sh

You can modify the configurations in lib/config.py to change the gpu_ids, network depth, image size, etc.

Testing

bash test.sh

Note

We release the AP-loss implementation in PyTorch instead of in MXNet due to an engineering issue: the python custom operator in MXNet does not run in parrallel when using multi-gpus. It is more practical to implement AP-loss in PyTorch, for faster training speed.

Acknowledgements

Many thanks to the pytorch implementation of RetinaNet at pytorch-retinanet.

Citation

If you find this repository useful in your research, please consider citing:

@inproceedings{chen2019towards,
  title={Towards accurate one-stage object detection with ap-loss},
  author={Chen, Kean and Li, Jianguo and Lin, Weiyao and See, John and Wang, Ji and Duan, Lingyu and Chen, Zhibo and He, Changwei and Zou, Junni},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={5119--5127},
  year={2019}
}