Orientated Silhouette Matching for Single-Shot Ship Instance Segmentation

Abstract

Object detection and semantic segmentation have achieved remarkable performance propelled by deep convolutional neural networks. However, neither of them can well parse and deal with swarms of rotating ships in remote sensing images. In this article, we pay more attention to the instance-level segmentation task, which recognizes objects more effectively and straightly. We propose a new network architecture, called orientated silhouette matching network, employing multiscale features and instance-level masks to enable single-shot and anchor-box-free instance segmentation. To be specific, we propose a novel-orientated polar template mask with orientated mask IoU to better match the ship silhouette. We also design a multiscale feature propagation and fusion module to improve the precision of detection. To further improve the performance, our network adopts Res2Net and Soft-NMS. Extensive experiments on the open datasets, namely Airbus Ship, demonstrate that our method improves the average precision by 14.2 and 10.0 percentage points on Res2Net101, compared with PolarMask and YOLACT. The source code will be open source after the reviewing process.