A large-scale multimodal dataset for perception in real-world maritime navigation scenarios

This dataset provides a large-scale multimodal perception benchmark for autonomous navigation in maritime environments. It consists of synchronized keyframes collected from real-world maritime scenes, with each frame including LiDAR point clouds, multi-view surround RGB images, ego-vehicle pose information, and velocity measurements.The dataset is annotated with instance-level 3D bounding boxes following the KITTI annotation format, covering multiple categories of maritime dynamic and static objects, including vessels (large, medium, and small), lighthouses, ball ranks, buoys, piers, reefs, and marine ranches. All annotations are provided with consistent object identities to support both 3D object detection and multi-object tracking tasks.In addition, structured natural language descriptions are included to facilitate high-level scene understanding for navigation and decision-making. This dataset is intended to support research in maritime perception, autonomous navigation, and multimodal scene understanding.

本数据集面向海事环境下的自主导航任务，提供大规模多模态感知基准测试集。其包含从真实海事场景采集的同步关键帧数据，每一帧均涵盖激光雷达（LiDAR）点云、多视角环绕RGB图像、自车位姿信息以及速度测量数据。该数据集采用KITTI标注格式，标注了实例级三维边界框，覆盖海事场景下多类动态与静态物体，包括大、中、小型船舶，灯塔、球标、浮标、码头、暗礁以及海洋牧场。所有标注均配有统一的物体身份标识，可同时支持三维目标检测与多目标跟踪任务。此外，数据集还包含结构化自然语言描述，以辅助完成导航与决策所需的高层场景理解任务。本数据集旨在为海事感知、自主导航以及多模态场景理解相关研究提供支撑。