A2D

该数据集名为A2D，包含了3782个视频，为演员和动作对提供了像素级别的密集标注，这些标注覆盖了稀疏采样帧的演员和动作的像素级别真值。该数据集的独特之处在于，它不仅为演员和动作类别提供了地面真实标注，还允许评估分割方法，这在像素级别标注能力方面是独一无二的。数据集的规模包括3782个视频，其中9561帧用于训练，2365帧用于测试，其任务是对演员-动作进行语义分割。

This dataset, named A2D, contains 3782 videos and provides pixel-level dense annotations for actor-action pairs, covering pixel-level ground truths of actors and actions in sparsely sampled frames. The unique feature of this dataset is that it not only offers ground-truth annotations for actor and action categories, but also enables the evaluation of segmentation methods, a capability unparalleled in the domain of pixel-level annotation tasks. Across these 3782 videos, 9561 frames are allocated for training and 2365 for testing, and its core task is semantic segmentation for actor-action pairs.