FPHA

我们收集了rgb-d视频序列，其中包含超过100K帧的45个日常手部动作类别，涉及多个手部配置中的26个不同对象。为了获得手部姿势注释，我们使用了自己的mo-cap系统，该系统通过6个磁传感器和逆运动学自动推断手部模型21个关节中每个关节的3D位置。此外，我们记录了6D对象的姿势，并为手-对象交互序列的子集提供了3D对象模型。据我们所知，这是第一个使用3D手部姿势研究第一人称手部动作的基准We collected RGB-D video sequences containing over 100K frames spanning 45 daily hand action categories, involving 26 distinct objects across multiple hand-object interaction configurations. To acquire hand pose annotations, we utilized our custom mo-cap system, which automatically infers the 3D positions of each of the 21 joints in the hand model via six magnetic sensors and inverse kinematics. Additionally, we recorded the 6D poses of the objects, and provided 3D object models for a subset of the hand-object interaction sequences. To the best of our knowledge, this is the first benchmark for investigating first-person hand actions using 3D hand poses.