SEAL

SEAL是一个大规模视频数据集，专注于多粒度时空动作定位。由OPPO研究院创建，该数据集包含两种类型的标注：SEAL Tubes和SEAL Clips。SEAL Tubes提供原子动作和复杂活动在tubelet级别的标注，涵盖49.6k原子动作和17.7k复杂活动。SEAL Clips则定位原子动作在空间和时间上的位置，生成510.4k动作标签。数据集的创建过程涉及从ActivityNet-1.3和HACS随机选择视频，并通过多对象跟踪、人物重新识别和标注者校正来获取tubelets。SEAL数据集的应用领域主要集中在视频理解，旨在解决未修剪视频中动作的精确空间位置和时间边界定位问题。

SEAL is a large-scale video dataset dedicated to multi-granularity spatio-temporal action localization. Developed by OPPO Research Institute, the dataset encompasses two annotation modalities: SEAL Tubes and SEAL Clips. SEAL Tubes provides tubelet-level annotations for both atomic actions and complex activities, covering 49.6k atomic action instances and 17.7k complex activity instances. SEAL Clips, on the other hand, localizes the spatial and temporal positions of atomic actions, producing 510.4k action labels. The construction of SEAL involves randomly sampling videos from ActivityNet-1.3 and HACS, followed by acquiring tubelets via multi-object tracking, person re-identification, and annotator correction. The primary application domain of the SEAL dataset is video understanding, where it aims to solve the problem of accurately localizing the spatial positions and temporal boundaries of actions in untrimmed videos.