This repository re-implements AC-FPN on the base of Detectron-Cascade-RCNN. Please follow Detectron on how to install and use this repo.
This repo has released CEM module without AM module, but we can get higher performance than the implementation of pytorch in paper. Also, thanks to the power of detectron, this repo is faster in training and inference.
The implementation of CEM is very simple, which is less than 200 lines code, but it can boost the performance almost 3% AP in FPN(resnet50).
AC-FPN can be readily plugged into existing FPN-based models and improve performance.
Visualization of object detection. Both models are built upon ResNet-50 on COCO minival.
Results of Mask R-CNN with (w) and without (w/o) our modules built upon ResNet-50 on COCO minival.
More detail in paper.
Because of the proposed architecture, We have better performance on most of FPN-base methods, especially on large objects.
The result of coco test-dev(team Neptune).
| backbone | type | lr schd |
im/ gpu |
box AP |
box AP50 |
box AP75 |
|---|---|---|---|---|---|---|
| X-152-32x8d-FPN-IN5k-baseline | Mask | s1x | 1 | 48.1 | 68.3 | 52.9 |
| X-152-32x8d-FPN-IN5k-cascade | Mask | s1x | 1 | 50.2 | 68.2 | 55.0 |
| X-152-32x8d-FPN-IN5k-acfpn(only CEM) | Mask | s1x | 1 | 51.9 | 70.4 | 57.0 |
If you use our code/model/data, please site our paper:
@article{cao2020attention,
title={Attention-guided Context Feature Pyramid Network for Object Detection},
author={Cao, Junxu and Chen, Qi and Guo, Jun and Shi, Ruichao},
journal={arXiv},
pages={arXiv--2005},
year={2020}
}
and Cascadercnn:
@inproceedings{cai18cascadercnn,
author = {Zhaowei Cai and Nuno Vasconcelos},
Title = {Cascade R-CNN: Delving into High Quality Object Detection},
booktitle = {CVPR},
Year = {2018}
}
and Detectron:
@misc{Detectron2018,
author = {Ross Girshick and Ilija Radosavovic and Georgia Gkioxari and
Piotr Doll\'{a}r and Kaiming He},
title = {Detectron},
howpublished = {\url{https://github.com/facebookresearch/detectron}},
year = {2018}
}