egocentric-origami-002

数据集名为egocentric-origami-002，是一个从第一人称视角视频中通过HaMeR技术提取的手部姿势数据集，采用LeRobotDataset v3.0格式封装。它记录人类手部动作（robot_type: human_hand），而非物理机器人。数据基于源视频origami_002.mp4重建，该视频使用iPhone、颈挂式支架和0.5倍广角镜头在家拍摄，覆盖折纸领域，包括三个任务：制作正方形基础、折叠鸟基础和完成折纸鹤。视频帧率为3fps，用于训练管道验证。数据集中右手检测率显著提高（episode 0: 99.6%, episode 1: 99.7%, episode 2: 100.0%）。注释通过Gemini API自动生成每帧叙述，但未经人工验证；时间轴在episode 2中进行了线性压缩校正。HaMeR估计了双手的3D关节位置，每手21个关节，总计126维特征。特征包括观察状态（126维）、动作（126维，表示下一帧的差值）和第一人称图像（224x224x3像素）。数据集统计信息：3个episodes、726帧、3个任务、3fps。The dataset named egocentric-origami-002 is a hand pose dataset extracted from egocentric (first-person perspective) videos using HaMeR technology and packaged in the LeRobotDataset v3.0 format. It records human hand movements (robot_type: human_hand) rather than physical robots. The data is reconstructed from the source video origami_002.mp4, which was shot at home using an iPhone, a neck-mounted stand, and a 0.5x wide-angle lens, covering the origami domain with three tasks: making a square base, folding a bird base, and completing an origami crane. The video frame rate is 3fps, used for training pipeline validation. The right-hand detection rate in the dataset shows significant improvement (episode 0: 99.6%, episode 1: 99.7%, episode 2: 100.0%). Annotations are automatically generated per-frame narratives via the Gemini API but are not human-verified; the timeline was linearly compressed and corrected in episode 2. HaMeR estimates 3D joint positions for both hands, with 21 joints per hand, totaling 126-dimensional features. Features include observation state (126 dimensions), action (126 dimensions, representing the difference to the next frame), and first-person images (224x224x3 pixels). Dataset statistics: 3 episodes, 726 frames, 3 tasks, 3fps.