Maritime vessel re-identification: novel VR-VCA dataset and a multi-branch architecture MVR-net

Abstract Maritime vessel re-identification (re-ID) is a computer vision task of identity matching across disjoint camera views. Prominent applications re-ID exist in the fields surveillance and maritime traffic flow analysis. However, field suffers from absence large-scale dataset that enables training deep learning models. In this study, we present new includes 4614 images 729 vessels along with 5-bin orientation 8-class vessel-type annotations to promote further research. A second contribution study baseline analysis our dataset. Performances 10 recent architectures are quantitatively compared reveal best practices. Lastly, propose novel multi-branch architecture, Vessel Re-ID network (MVR-net), address challenging problem re-ID. Evaluation approach on yields 74.5% mAP 77.9% Rank-1 score, providing performance increase 5.7% 5.0% over best-performing baseline. MVR-net also outperforms PRN (a pioneering vehicle network), by 2.9% 4.3% higher Rank-1, respectively.