FGSR: A Fine-Grained Ship Retrieval Dataset and Method in Smart Cities

Abstract

Ship reidentification is an important part of water transportation systems in smart cities. Existing ship reidentification methods lack a large-scale fine-grained ship retrieval dataset in the wild and existing ship recognition solutions mainly focus on the ship target identification rather than the fine-grained ship reidentification. Furthermore, previous ship target identification systems are usually based on synthetic aperture radar (SAR) image, automatic identification system (AIS) data, or video streaming, which is confronted with expensive deployment costs, such as the installation cost of SAR and AIS, and the communication and storage overhead. Indeed, ship reidentification benefits for traffic monitoring, navigation safety, vessel tracking, etc. To address these problems, we propose a new large-scale fine-grained ship retrieval dataset (named FGSR) that consists of 30,000 images of 1000 ships captured in the wild. Besides, to tackle the difficulty of spatial-temporal inconsistency in ship identification in the wild, we design a multioriented ship reidentification network named FGSR-Net that consists of three modules to address different crucial problems. The pyramid fusion module was aimed at addressing the problem of variant size and shape of ship targets, the occlusion modules attempt to detect the unchangeable area of ship images, while the multibranch identity module generates discriminative feature representation for ship targets from different orientations. Experimental evaluations on FGSR dataset show the effectiveness and efficiency of our proposed FGSR-Net. The mean average precision of ship reidentification is around 92.4%, and our FGSR-Net proposed method only takes 3 seconds to give the retrieval results from 30,000 images.