First-Person Hand Action Benchmark

本数据集名为‘First-Person Hand Action Benchmark’，由帝国理工学院创建，包含超过100,000帧的RGB-D视频序列，涵盖45种日常手部动作类别，涉及26种不同物体。数据集通过磁性传感器和逆向运动学自动获取手部21个关节的3D位置，并提供部分物体的6D姿态和3D模型。创建过程涉及多种手部配置和时间动态的覆盖。该数据集适用于3D手部姿态估计、6D物体姿态及机器人学等领域，旨在解决手部动作识别中的精确度问题。

This dataset, named 'First-Person Hand Action Benchmark', was created by Imperial College London. It contains over 100,000 RGB-D video frame sequences, covering 45 categories of daily hand actions and involving 26 distinct objects. The dataset automatically acquires the 3D positions of 21 hand joints via magnetic sensors and inverse kinematics, and provides 6D poses and 3D models for a subset of the included objects. Its creation process covers diverse hand configurations and temporal dynamics. This benchmark is applicable to research fields such as 3D hand pose estimation, 6D object pose estimation, and robotics, aiming to address the accuracy issues in hand action recognition.