Introduction

This repository implements some deep-learning models for ship detection on SAR images. All models are developed to work with the Ship SAR Detection Dataset (SSDD).

Moreover, the following work is based on the OpenMMLab Detection Toolbox and Benchmark.

Installation

Step 1: create and activate conda environment with python 3.8

conda create -n myenv python=3.8
conda activate myenv

Step 2: install MMCV using MIM.

pip install -U openmim
mim install mmcv-full

Step 3: install project from source

git clone https://github.com/ALLARDLE/ShipSARDetect_mmdetection.git
cd ShipSARDetect_mmdetection
pip install -v -e .

Step 4: download Ship SAR Detection Dataset in data/ folder

Getting Started

Base configuration of SSDD is stored in config/_base_/datasets/ssdd_detection.py All implemented model for SSDD are stored in config/ssdd/ folder.

Train models

Faster R-CNN:

• Run Faster R-CNN VGG16: python tools\train.py configs\ssdd\faster_rcnn_vgg16_fpn_ssdd.py
• Run Faster R-CNN ResNet50: python tools\train.py configs\ssdd\faster_rcnn_r50_fpn_ssdd.py
• Run Faster R-CNN ResNet50 pretrained: python tools\train.py configs\ssdd\faster_rcnn_r50_fpn_ssdd_pretrained.py

Cascade R-CNN:

• Run Cascade R-CNN VGG16: python tools\train.py configs\ssdd\cascade_rcnn_vgg16_fpn_ssdd.py
• Run Cascade R-CNN ResNet50: python tools\train.py configs\ssdd\cascade_rcnn_r50_fpn_ssdd.py
• Run Cascade R-CNN Swin: python tools\train.py configs\ssdd\cascade_rcnn_swin_fpn_ssdd.py
• Run Cascade R-CNN Swin pretrained: python tools\train.py configs\ssdd\cascade_rcnn_swin_fpn_ssdd_pretrained.py

Test models

python tools\test.py configs\ssdd\faster_rcnn_r50_fpn_ssdd.py work_dirs\faster_rcnn_r50_fpn_ssdd\latest.pth --show-dir results

In development

In order to implement ESTDNet model from Ship Detection in SAR Images Based on Feature Enhancement Swin Transformer and Adjacent Feature Fusion article some changes in mmdet library.

Backbone: added Feature Enhancement Swin module as feswin.py and feswinv2.py. v2 is based on Swin module of MMDetection whereas the v1 is based on PyTorch one.

Neck: added Adjacent Feature Fusion module as aff.py

License

This project is released under the Apache 2.0 license.