Multitask learning for ship detection from synthetic aperture radar images

This is the official implementation of MTL-Det (JSTARs), a SAR ship detection method. For more details, please refer to:

Multitask learning for ship detection from synthetic aperture radar images [Paper]
Xin Zhang , Chunlei Huo , Nuo Xu , Hangzhi Jiang, Yong Cao, Lei Ni and Chunhong Pan

Installation

Please refer to install.md for installation.

Getting Started

Preparation

Clone the code

git clone https://github.com/XinZhangNLPR/JSTARs_MTLDet.git

Download the model weight used in the paper:

HRSID dataset

	Backbone	AP	AP@50	AP@75	AP_S	AP_M	AP_L	download
MTL-Det	ResNeXt-101-64×4	68.0	89.5	77.7	68.7	69.6	25.8	Google

Put the model to work_dirs/HTL_1x_renext/

LSSDD-v1.0 dataset

	Backbone	Off-shore	In-shore	ALL	download
MTL-Det	ResNet-50	88.7	38.7	71.7	Google

Put the model to work_dirs/HTL_1x_faster/

Evaluate

1.Multi-GPUs Test

./tools/dist_test.sh work_dirs/HTL_1x_faster/HTL_ins_faster_rcnn_r50_fpn_1x_hrsid.py work_dirs/HTL_1x_faster/epoch_11.pth 8 --eval mAP

2.Single-GPU Test

python tools/test.py work_dirs/HTL_1x_faster/HTL_ins_faster_rcnn_r50_fpn_1x_hrsid.py work_dirs/HTL_1x_faster/epoch_11.pth --eval mAP

Citation

@article{zhang2021multitask,
  title={Multitask learning for ship detection from synthetic aperture radar images},
  author={Zhang, Xin and Huo, Chunlei and Xu, Nuo and Jiang, Hangzhi and Cao, Yong and Ni, Lei and Pan, Chunhong},
  journal={IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing},
  volume={14},
  pages={8048--8062},
  year={2021},
  publisher={IEEE}
}