fhswf/COCO-WB133-Quaternion-Mapping-Dataset

WholeBody3D-to-AvatarPose数据集是一个合成数据集，旨在学习从3D全身姿势估计输出到以四元数表示的虚拟角色关节旋转的映射。其主要目标是训练能够将COCO WholeBody（133个关键点）3D关节坐标转换为物理上合理的虚拟角色姿势的模型，适用于实时动画、手语虚拟角色和具身代理。数据集包含无噪声的3D关节位置和精确的地面真实关节旋转，通过Blender合成生成。数据格式包括JSON文件，包含元数据和骨骼对象，详细描述了骨骼的局部、骨架空间和世界空间变换。数据集适用于虚拟角色动画、手语虚拟角色合成、具身AI、运动重定向和基于四元数的姿势回归等研究和开发。

The WholeBody3D-to-AvatarPose Dataset is a synthetic dataset designed to learn a mapping from 3D whole-body pose estimation outputs to avatar joint rotations represented as quaternions. The primary objective is to enable the training of models that convert COCO WholeBody (133 keypoints) 3D joint coordinates into physically plausible avatar poses, suitable for real-time animation, sign language avatars, and embodied agents. The dataset includes noise-free 3D joint positions and exact ground truth joint rotations, synthetically generated in Blender. The data format consists of JSON files with metadata and bone objects, detailing local, armature space, and world space transformations. The dataset is intended for research and development in avatar animation, sign language avatar synthesis, embodied AI, motion retargeting, and quaternion-based pose regression.