AFPN: Asymptotic Feature Pyramid Network for Object Detection (arXiv)
By Guoyu Yang, Jie Lei, Zhikuan Zhu, Siyu Cheng, Zunlei Feng, Ronghua Liang
This project is based on mmdetection.
mmengine==0.7.3
mmcv==2.0.0
mmdet==3.0.0
Please refer to mmdetection for installation.
AFPN
├── mmdetection
├── data
│ ├── coco
│ │ ├── annotations
│ │ ├── train2017
│ │ ├── val2017
│ │ ├── test2017
├── faster-rcnn_r50_afpn_1x_coco.py
├── train.py
├── test.py
Single gpu for train:
CUDA_VISIBLE_DEVICES=0 python ./mmdetection/tools/train.py faster-rcnn_r50_afpn_1x_coco.py --work-dir ./weight/
Multiple gpus for train:
CUDA_VISIBLE_DEVICES=0,1 ./mmdetection/tools/dist_train.sh faster-rcnn_r50_afpn_1x_coco.py 2 --work-dir ./weight/
Train in pycharm: If you want to train in pycharm, you can run it in train.py.
see more details at mmdetection.
CUDA_VISIBLE_DEVICES=0 _DEVICES=1 python ./mmdetection/tools/test.py faster-rcnn_r50_afpn_1x_coco.py <CHECKPOINT_FILE>
For example,
CUDA_VISIBLE_DEVICES=0 python ./mmdetection/tools/test.py faster-rcnn_r50_afpn_1x_coco.py ./weight/afpn_weight.pth
Test in pycharm: If you want to test in pycharm, you can run it in test.py.
see more details at mmdetection.
Detector | Backbone | Image size | GFLOPs | Params (M) | AP | AP0.5 | AP0.75 | Weight |
---|---|---|---|---|---|---|---|---|
Faster R-CNN + FPN | ResNet-50 | 640 x 640 | 91.3 | 41.8 | 37.4 | 57.3 | 40.3 | None |
Faster R-CNN + AFPN | ResNet-50 | 640 x 640 | 89.7 | 49.8 | 39.0 | 57.6 | 42.0 | Link |
If you find AFPN useful in your research, please consider citing:
@article{yang2023afpn,
title={AFPN: Asymptotic Feature Pyramid Network for Object Detection},
author={Yang, Guoyu and Lei, Jie and Zhu, Zhikuan and Cheng, Siyu and Feng, Zunlei and Liang, Ronghua},
journal={arXiv preprint arXiv:2306.15988},
year={2023}
}