VesselNet: A Large-Scale Dataset and Efficient Mixed Attention Network for Vessel Re-identification
Zijun Yu, Jin Liu, Shenjie Zou, Yuetian Cao · 2023
Compared to the rapid development of pedestrian and vehicle re-identification in urban safety management, many challenging issues of vessel re-identification have not been fully investigated. Such as more dramatic viewpoint changes, more extensive lighting conditions (reflected light from the sea), and more complex navigational backgrounds. To facilitate research on vessel Re-ID at sea, we have collected a new dataset called VeRiS, which has the following salient features: 1) the original images of ships are captured by professional photographers from different times and places around the world via airborne or land-based cameras; 2) the images have been processed with fine detection to remove redundant and complex backgrounds; 3) the dataset is huge in size and has a large number of ship types. VeRiS contains 150,623 images of 2,904 ship IDs. Furthermore, we propose a novel vessel Re-ID network, named VesselNet, that places greater emphasis on the ships in the images. Specifically, we incorporate an improved hybrid attention module into the network, which alters the original pooling method to enhance the differentiation of distinct features. Experiments indicate that our VesselNet performs favorably on both the proposed VeRiS and another vehicle dataset.