RoMoDataset/RoMo-SMPL
收藏资源简介:
RoMo-SMPL是一个基于RoMo论文的大规模人体运动数据集,专注于SMPL身体参数空间(包括全局方向、21个关节的身体姿势、形状和平移参数)。该数据集包含815,555个运动序列,分为训练集(693,607)、验证集(81,270)和测试集(40,678),帧率为30 fps。身体模型来源于GVHMR恢复的SMPL身体。每个运动片段配有五个文本描述(从标签到段落级别)和一个三级语义分类法(包括类别、子类别和原子动作标签)。数据以Parquet格式(zstd分片)存储,许可证为CC BY-NC 4.0。该数据集主要用于文本到3D、文本生成和人体运动生成等任务,但不包含手部姿势和手部描述。
RoMo-SMPL is a paper-aligned release of the RoMo body motion corpus in SMPL body parameter space (global orientation, 21-joint body pose, shape, translation). It contains 815,555 sequences with fixed train (693,607), val (81,270), and test (40,678) splits, at a frame rate of 30 fps. The body model is derived from GVHMR recovery of SMPL body. Each clip includes five text captions and a three-level semantic taxonomy (category, subcategory, atomic action). The data is stored in Parquet format (zstd shards) under the CC BY-NC 4.0 license. It is designed for tasks such as text-to-3D, text generation, and motion generation, but does not include hand pose or hand captions.




