ZhanQiQi/AugFPN

AugFPN: Improving Multi-scale Feature Learning for Object Detection(CVPR 2020)

By Chaoxu Guo, Bin Fan, Qian Zhang, Shiming Xiang, Chunhong Pan

arxiv paper, pdf

This project is based on mmdetection

Introduction

Current state-of-the-art detectors typically exploit feature pyramid to detect objects at different scales. Among them, FPN is one of the representative works that build a feature pyramid by multi-scale features summation. However, the design defects behind prevent the multi-scale features from being fully exploited. In this paper, we begin by first analyzing the design defects of feature pyramid in FPN, and then introduce a new feature pyramid architecture named AugFPN to address these problems. Specifically, AugFPN consists of three components: Consistent Supervision, Residual Feature Augmentation, and Soft RoI Selection. AugFPN narrows the semantic gaps between features of different scales before feature fusion through Consistent Supervision. In feature fusion, ratio-invariant context information is extracted by Residual Feature Augmentation to reduce the information loss of feature map at the highest pyramid level. Finally, Soft RoI Selection is employed to learn a better RoI feature adaptively after feature fusion.

Install

Please refer to INSTALL.md for installation.

Prepare data

  mkdir -p data/coco
  ln -s /path_to_coco_dataset/annotations data/coco/annotations
  ln -s /path_to_coco_dataset/train2017 data/coco/train2017
  ln -s /path_to_coco_dataset/test2017 data/coco/test2017
  ln -s /path_to_coco_dataset/val2017 data/coco/val2017

Pretrained Models

Pretrained models will be available.

Training

./tools/dist_train.sh <CONFIG_FILE> <GPU_NUM> --validate --work_dir <WORK_DIR>

For example,

./tools/dist_train.sh configs/faster_rcnn_r50_augfpn_1x.py 8 --validate --work_dir faster_rcnn_r50_augfpn_1x

see more details at mmdetection

Testing

python tools/test.py <CONFIG_FILE> <CHECKPOINT_FILE> --gpus <GPU_NUM> --out <OUT_FILE> --eval <EVAL_TYPE>

When test results of detection, use --eval bbox. When test results of instance segmentation, use --eval bbox segm. See more details at mmdetection.

For example,

python tools/test.py configs/mask_rcnn_r50_augfpn_1x.py <CHECKPOINT_FILE> --gpus 8 --out results.pkl --eval bbox segm

Results on MS COCO testdev2017

Backbone	detector	mAP(mask)	mAP(det)
ResNet-50 FPN	Faster R-CNN	-	36.5
ResNet-50 AugFPN	Faster R-CNN	-	38.8
ResNet-50 FPN	Mask R-CNN	34.4	37.5
ResNet-50 AugFPN	Mask R-CNN	36.3	39.5
ResNet-50 FPN	RetinaNet	-	35.9
ResNet-50 AugFPN	RetinaNet	-	37.5
ResNet-50 FPN	FCOS	-	37.0
ResNet-50 AugFPN	FCOS	-	37.9

Citations

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

@misc{guo2019augfpn,
    title={AugFPN: Improving Multi-scale Feature Learning for Object Detection},
    author={Chaoxu Guo and Bin Fan and Qian Zhang and Shiming Xiang and Chunhong Pan},
    year={2019},
    eprint={1912.05384},
    archivePrefix={arXiv},
    primaryClass={cs.CV}
}

License

This project is released under the Apache 2.0 license

ZhanQiQi / AugFPN