GOT-10k
收藏资源简介:
GOT-10k是一个大规模、高多样性的基准数据集,用于在野外进行通用目标跟踪。该数据集由中国科学院自动化研究所创建,包含超过10,000个视频片段和150万个手动标注的边界框。GOT-10k是首个使用WordNet语义层次结构指导类别填充的视频轨迹数据集,确保了移动目标的全面和相对无偏的覆盖。此外,GOT-10k首次引入了单次跟踪器评估协议,训练和测试类别完全不重叠,避免了偏向熟悉对象的评估结果,促进了跟踪器开发中的泛化能力。数据集还提供了额外的标签,如运动类别和目标可见比率,便于开发对运动和遮挡敏感的跟踪器。GOT-10k的应用领域包括监控、增强现实、生物学和机器人技术,旨在解决在无约束环境中跟踪移动目标的挑战。
GOT-10k is a large-scale, high-diversity benchmark dataset for generic object tracking in the wild. Developed by the Institute of Automation, Chinese Academy of Sciences, it contains over 10,000 video clips and 1.5 million manually annotated bounding boxes. As the first video tracking dataset that uses the WordNet semantic hierarchy to guide category filling, GOT-10k ensures comprehensive and relatively unbiased coverage of moving objects. Additionally, it introduces the one-shot tracker evaluation protocol for the first time, where training and test categories are completely non-overlapping, thus avoiding evaluation results biased towards familiar objects and promoting generalization capability in tracker development. The dataset also provides supplementary labels including motion categories and target visibility ratios, which facilitate the development of trackers sensitive to motion and occlusion. Its application scenarios cover surveillance, augmented reality, biology and robotics, with the goal of addressing the challenges of tracking moving objects in unconstrained environments.




