Branch Augmented Neural Network for Marine Vessel Re-Identification Using LiDAR Point Cloud

This paper introduces the first marine vessel re-identification model using 3D LiDAR point clouds. Although various models for processing 3D LiDAR data have been developed, their application to marine vessel re-identification remains underexplored. In this work, we evaluate the performance of multiple existing point cloud processing models specifically for vessel re-identification tasks and propose a novel model, the Branch-Augmented Neural Network (BANN), designed to effectively aggregate multiple feature representations and capture diverse aspects of vessel geometry and appearance. For evaluation purposes, we introduce the Jeongok-Vessels dataset, comprising 360-degree point cloud views of eight distinct vessels. This dataset was collected at Jeongok Port, Republic of Korea, using a camera-to-LiDAR back-projection technique. Experimental results validate the feasibility of 3D LiDAR point cloud-based marine vessel re-identification, demonstrating that BANN's multi-branch feature aggregation significantly enhances re-identification accuracy compared to baseline methods.