A2D (Actor-Action Dataset)

人类会飞吗？坚决不。汽车能吃吗？再次，绝对不是。然而，这些荒谬的推论是由于当前对行动理解中特定类型的行为者的漠视造成的。我们不知道同时推断视频中的演员和动作的工作，更不用说要试验的数据集了。因此，A2D 标志着计算机视觉社区首次努力共同考虑各种类型的参与者进行各种行动。确切地说，我们考虑了七个演员类别（成人、婴儿、球、鸟、汽车、猫和狗）和八个动作类别（爬、爬、吃、飞、跳、滚、跑和走），不包括无动作类，我们也考虑。 A2D 有 3782 个视频，每个有效的 actor-action 元组至少有 99 个实例，并且视频标记有像素级 actor 和采样帧的动作。 A2D 数据集可作为各种视觉问题的新型大规模测试平台：视频级单标签和多标签演员动作识别、实例级对象分割/协同分割以及像素级演员动作语义细分仅举几例。Can humans fly? Absolutely not. Can cars be eaten? Once again, definitely not. However, these absurd inferences stem from the neglect of specific types of actors in current action understanding research. Prior to this work, there has been no research focusing on jointly inferring actors and actions in videos, let alone a dedicated dataset for such studies. Therefore, A2D marks the first endeavor by the computer vision community to jointly consider diverse actors performing a wide range of actions. Specifically, we consider seven actor categories (adults, infants, balls, birds, cars, cats, and dogs) and eight action categories (climb, climb, eat, fly, jump, roll, run, walk), and we also take the no-action class into account, even though it was not included in the initial listing. A2D contains 3782 videos, with each valid actor-action tuple having at least 99 instances, and all videos are annotated with pixel-level actors and actions for sampled frames. The A2D dataset serves as a novel large-scale testbed for various computer vision tasks: video-level single-label and multi-label actor-action recognition, instance-level object segmentation/cosegmentation, and pixel-level actor-action semantic segmentation, to name just a few.

搜集汇总

背景与挑战

背景概述

A2D数据集是计算机视觉领域首个共同考虑视频中多种类型演员和动作的数据集，包含3782个视频，覆盖七个演员类别和八个动作类别，每个有效的actor-action元组至少有99个实例，并标记有像素级演员和采样帧的动作。它可作为视频级识别、实例级分割和像素级语义分割等视觉问题的测试平台，由密歇根大学于2015年发布。