WakeNet

A CNN-based optical image ship wake detector. It uses SWIM as the benchmark dataset.

If you find this code helpful in your research, please cite this paper:

@ARTICLE{9618934,
author={Xue, Fuduo and Jin, Weiqi and Qiu, Su and Yang, Jie},
journal={IEEE Transactions on Geoscience and Remote Sensing},
title={Rethinking Automatic Ship Wake Detection: State-of-the-Art CNN-based Wake Detection via Optical Images},
year={2021},
volume={},
number={},
pages={1-1},
doi={10.1109/TGRS.2021.3128989}}

Introduction

Most existing wake detection algorithms use Radon transform (RT) due to the long streak features of ship wakes in SAR images. The high false alarm rate of RT requires that the algorithms have significant human intervention for image preprocessing. When processing optical images these algorithms are greatly challenged because it is not only sea clutter that interferes with the wake detection in the optical waveband, but many other environmental factors. To solve the problem, in this repo, we address the automatic ship wake detection task in optical images from the idea of the CNN to design an end-to end detector as a novel method. The detector processes all the wake textures clamped by the V-shaped Kelvin arms as the object, and conducts detection via the OBB. The extra regression of wake tip coordinates and Kelvin arm direction can improve the hard wake detection performance while predicting the wake heading. Some special structures and strategies are also adopted according to the wake features.

Detection Results on SWIM dataset

Method	Baseline	Backbone	mAP (AP)
R²CNN	Faster R-CNN	ResNet-101	65.24%
R-RetinaNet	RetinaNet	ResNet-101	72.22%
R-YOLOv3	YOLOv3	DarkNet-53	54.87%
BBAVectors	CenterNet	ResNet-101	66.19%
WakeNet (Ours)	RetinaNet	FcaNet-101	77.04%

Dependencies

Python 3.6.12, Pytorch 1.7.0, OpenCV-Python 3.4.2, CUDA 11.0

Getting Started

Installation

Build modules for calculating OBB overlap and performing Radon transform:

cd $ROOT/utils
sh make.sh

Prepare the ship wake dataset

WakeNet uses SWIM dataset collected by us as the benchmark dataset. The images in SWIM dataset are uniformly cropped into the size of 768 × 768, with 6,960 images for training, 2,320 images for testing, and 2,320 images for evaluation.

Download SWIM dataset from Kaggle into the $ROOT directory.
Modify the dataset path in datasets/generate_imageset.py and then run it.
Run datasets/evaluate/swim2gt.py to generate ground truth labels for evaluation.

You may modify the above two files to support other datasets.

Train model

You can modify the training hyperparameters with the default config file hyp.py. Then, run the following command.

python train.py

Evaluate model

Test the mAP on the SWIM dataset. Note that the VOC07 metric is defaultly used.

python eval.py

Test model

Generate visualization results.

python demo.py

Code reference

Thanks to Rotated-RetinaNet for providing a reference in our code writing.

Lilytopia/WakeNet