inclusionAI/OpenAoE-2000h
收藏资源简介:
Open-AoE是一个以自我为中心(第一人称)视角的手部操作数据集,包含3D手部重建数据、相机轨迹和细粒度原子动作注释。数据来源于真实世界智能手机拍摄的视频,经过基于视觉SLAM的相机轨迹估计和参数化手部重建(使用MANO模型)处理。数据集提供完整的MANO 3D手部姿态参数(包括15个手指关节姿态和10维形状参数)、度量尺度的相机轨迹(世界坐标系和相机坐标系下的变换矩阵),以及细粒度的原子动作注释(分段级别,包括场景、手部归属和动作描述)。数据规格包括30 FPS帧率、1280×720或1920×1080分辨率。数据集还包含原始视频、去失真视频、相机参数文件(如video_info.json和undistorted_video_info.json)和可视化结果。该数据集适用于手部姿态估计、动作识别、增强现实和机器人操作等研究。
Open-AoE is an egocentric (first-person) hand manipulation dataset containing 3D hand reconstruction data, camera trajectories, and fine-grained atomic action annotations. The data originates from real-world smartphone-captured videos, processed via a visual-SLAM-based camera trajectory estimation and parametric hand reconstruction (using the MANO model). The dataset provides complete MANO 3D hand pose parameters (including 15 finger joint poses and 10-dimensional shape parameters), metric-scale camera trajectories (transformation matrices in world and camera coordinate systems), and fine-grained atomic action annotations (segment-level, including scene, hand attribution, and action descriptions). Specifications include 30 FPS frame rate and 1280×720 or 1920×1080 resolution. The dataset also includes raw videos, undistorted videos, camera parameter files (e.g., video_info.json and undistorted_video_info.json), and visualization results. It is suitable for research in hand pose estimation, action recognition, augmented reality, and robotic manipulation.




