Optical-to-SAR Domain Adaptive with Inversion Regularization for Unsupervised Ship Detection

Introduction

This is a PyTorch/GPU implementation of the unpublished paper Optical-to-SAR Domain Adaptive with Inversion Regularization for Unsupervised Ship Detection, implemented by shijie Wang(wang.shijie@stu.xjtu.edu.cn).

This implementation is built on Domain Adaptive Faster R-CNN. If you find this repository useful, please cite the oringinal paper:

@inproceedings{chen2018domain,
  title={Domain Adaptive Faster R-CNN for Object Detection in the Wild},
      author =     {Chen, Yuhua and Li, Wen and Sakaridis, Christos and Dai, Dengxin and Van Gool, Luc},
      booktitle =  {Computer Vision and Pattern Recognition (CVPR)},
      year =       {2018}
  }

@misc{massa2018mrcnn,
author = {Massa, Francisco and Girshick, Ross},
title = {{maskrnn-benchmark: Fast, modular reference implementation of Instance Segmentation and Object Detection algorithms in PyTorch}},
year = {2018},
howpublished = {\url{https://github.com/facebookresearch/maskrcnn-benchmark}},
note = {Accessed: [Insert date here]}
}

Installation

Install the corresponding package according to requirements.txt and use DAIR.

Example Usage

An example of DAIR adapting from DIOR dataset to SSDD dataset is provided:

1. Please download dataset DIOR and SSDD.
2. Symlink the path to the DIOR and SSDD dataset to datasets/ as follows:

   # symlink the dataset
   cd ~/DAIR
   ln -s /<path_to_DIOR_dataset>/ datasets/DIOR
   ln -s /<path_to_SSDD_dataset>/ datasets/SSDD

3. Ensure that ~/DAIR/maskrcnn_benchmark/modeling/reverse_backbone is in your system path. The training log and weight files are saved in the output directory, and you can also update the output location through ~/DAIR/maskrcnn_benchmark/config/defaults
4. Train the DAIR:

   python tools/train_net.py --config-file "configs/DAIR_sar/e2e_da_faster_rcnn_R_50_C4_SARdior_to_sar.yaml" 
   

5. Test the trained model:

   python tools/test_net.py --config-file "configs/DAIR_sar/e2e_da_faster_rcnn_R_50_C4_SARdior_to_sar.yaml" 
   

Pretrained Model & Results

~Pretrained model with DAIR on Resnet-50 bakcbone for DIOR->SSDD task is provided.

The following results are all tested with Resnet-50 backbone.

	IRM	MLF	TCE	AP@50	AP@75	APm
DA Faster R-CNN				67.9	17.0	29.1
DAIR		✓	✓	66.5	26.2	33.0
DAIR	✓		✓	69.8	18.8	30.3
DAIR	✓	✓		65.7	30.1	32.9
DAIR	✓	✓	✓	73.0	33.0	37.0