Kinetics-TPS

We develop a large-scale Kinetics-Temporal Part State (Kinetics-TPS) benchmark for this study. Different from existing video action datasets, our Kinetics-TPS provides 7.9M annotations of 10 body parts, 7.9M part state (i.e., how a body part moves) and 0.5M interactive objects in the video frames of 24 human action classes, which bring new opportunity to understand human action by compositional learning of body parts.

搜集汇总

背景与挑战

背景概述

Kinetics-TPS是一个基于中层语义标注的人体行为数据库，专注于人体部位级别的动作分析。该数据集从Kinetics-700中选取了24个行为类别，包含4741个视频，并提供了丰富的人体、部位和物体标注，以支持对复杂行为的深入理解。