LEDet: localization estimation detector with data augmentation for ship detection based on unmanned surface vehicle

Abstract

Ship detection is significant for monitoring ports, especially contributing to the safe driving of Unmanned Surface Vehicle (USV). However, recent ship detection based on deep learning lacks complete ship datasets and uses the classification score as the ranking basis, which harms their performance. To address the problems, we present a one-stage localization estimation detector (LEDet) with ship-customized data augmentation. Specifically, we integrate the localization quality estimation into the classification branch as a soft label localization score. We further apply ship-customized data augmentation named “cutting-transform-paste” to expand ship datasets without manual annotation. Hence, a large number of diverse ship datasets can be created. Extensive experiments show that our LEDet consistently exceeds the strong baseline by 8.0% COCO-style Average Precision (AP) with ResNet-50. It significantly improves the ship detection performance.