Cross-Scale Attention Feature Pyramid Network for Challenge Underwater Object Detection

Miao yang, Jinyang Zhong, Hansen Zhang, Can Pan, Xinmiao Gao, Chenglong Gong

Abstract

Underwater object detection (UOD) is more difficult than common object detection tasks due to the noise caused by irrelevant objects and textures, and the scale variation. These difficulties pose a higher challenge to the feature extraction capability of the detectors. Feature pyramid network enhances the scale detection capability of detectors, while attention mechanisms effectively suppress irrelevant features. We present here a cross-scale attention feature pyramid network (CSAFPN) for UOD. A feature fusion guided (FFG) module is incorporated in the CSAFPN, which constructs cross-scale context information and simultaneously guides the enhancement of all feature maps. Compared to existing FPN-like architectures, CSAFPN excels not only in capturing cross-scale long-range dependencies but also in acquiring compact multi-scale feature maps that specifically emphasize target regions. Extensive experiments on the Brackish2019 dataset show that CSAFPN can achieve consistent improvements on various backbones and detectors. Moreover, FFG can be seamlessly integrated into any FPN-like architecture, offering a cost-effective improvement in UOD, resulting in a 1.4% AP increase for FPN, a 1.3% AP increase for PANet, and a 1.4% AP increase for NAS-FPN.

Dependencies

• Python==3.7.12
• PyTorch==1.10.0+cu111
• mmdetection==2.4.0
• mmcv==1.1.5

Installation

The basic installation follows with mmdetection. It is recommended to use manual installation.

Datasets

Brackish download link: BaiduYun

The structure of this dataset is:

├── Brackish dataset
│   ├── annotations
│   │   ├── train.json
│   │   ├── val.json
│   │   ├── test.json
│   ├── images
│   │   ├── train
│   │   ├── val
│   │   ├── test

Training

python tools/train.py work_dir/csafpn1/optics.py

Testing

python tools/test.py work_dir/csafpn1/optics.py <path/to/checkpoints> --eval bbox

Results on different detector

Backbone	Neck	detector	mAP	AP50	AP75
ResNet-50	FPN	Faster R-CNN	76.9	96.8	87.9
ResNet-50	CSAFPN	Faster R-CNN	77.9	97.4	87.8
ResNet-50	FPN	Mask R-CNN	76.8	96.7	87.8
ResNet-50	CSAFPN	Mask R-CNN	78.8	98.2	89.2
ResNet-50	FPN	Cascade R-CNN	81.1	98.0	90.8
ResNet-50	CSAFPN	Cascade R-CNN	82.5	98.2	92.8

Results on different backbone

ne	Neck	detector	mAP	AP50	AP75
ResNet 18	FPN	Cascade R-CNN	74.5	95.2	83.1
ResNet 18	CSAFPN	Cascade R-CNN	76.0	95.9	86.0
ResNet 50	FPN	Cascade R-CNN	81.1	98.0	90.8
ResNet 50	CSAFPN	Cascade R-CNN	82.5	98.2	92.8
ResNet 101	FPN	Cascade R-CNN	81.0	97.8	91.6
ResNet 101	CSAFPN	Cascade R-CNN	82.2	98.2	92.2
ResNeXt50_32×4d	FPN	Cascade R-CNN	74.0	95.4	83.8
ResNeXt50_32×4d	CSAFPN	Cascade R-CNN	74.6	95.8	84.5

License

This project is released under the Apache 2.0 license