MVTD

Tracking objects in maritime environments remains a challenging task due to dynamic water surfaces, occlusions, reflections, and camera-induced motion blur. To advance research in this domain, we introduce MVTD (Maritime Visual Tracking Dataset), a comprehensive benchmark comprising 182 sequences and over 150,000 frames, each annotated with precise bounding boxes. The dataset captures a wide variety of real-world maritime conditions and includes annotations for 9 critical attributes, such as occlusion, low resolution, background clutter, and motion blur, enabling fine-grained performance analysis. Additionally, MVTD includes four object types commonly encountered in maritime surveillance scenarios. To evaluate current capabilities, we benchmark 14 SOTA trackers on MVTD, revealing notable performance limitations and motivating the development of domain specific tracking solutions for maritime environments.在海事环境中开展目标跟踪始终是一项极具挑战性的任务，这是由于动态水面、目标遮挡、反射效应以及相机引入的运动模糊等因素所致。为推动该领域的研究发展，我们构建了海事视觉跟踪数据集（MVTD，Maritime Visual Tracking Dataset），这是一个涵盖182个序列与超15万帧图像的综合性基准测试集，所有数据均配有精确的边界框标注。该数据集收录了丰富多样的真实海事场景条件，并为9项关键属性提供标注，包括遮挡、低分辨率、背景杂乱以及运动模糊等，可支持对跟踪模型进行细粒度的性能分析。此外，该数据集涵盖了海事监视场景中常见的四类目标类型。为评估现有跟踪算法的实际性能，我们在该数据集上对14个当前最优（SOTA）跟踪器开展了基准测试，结果揭示了现有算法存在的显著性能局限，同时也推动了海事环境领域专用跟踪解决方案的研发。