RoMoDataset/RoMo-HML-272
收藏资源简介:
RoMo-HML-272 是一个大规模人体运动数据集,采用272维运动特征表示,基于HumanML3D风格编码(263维标准特征加9维绝对3D/接触增强特征)。数据集包含815,555个运动序列,以30帧每秒的帧率存储,格式为Parquet(zstd压缩分片)。每个序列配有5个级别的文本描述(从标签到段落),由视觉语言模型生成。数据集目前为私有/内部使用,将在RoMo论文发布时公开。许可为CC BY-NC 4.0。该数据集适用于文本到3D、运动生成等任务,并提供了从特征解码到关节位置的工具。
RoMo-HML-272 is a large-scale human body motion dataset packed in a 272-dimensional motion-feature representation that augments the standard 263-D HumanML3D encoding with an additional 9 dimensions for absolute-3D/contact-augmented features. It contains 815,555 sequences at 30 fps, stored in Parquet format (zstd-compressed shards). Each sequence has 5 levels of text descriptions (from tag to paragraph) generated by a vision-language model. The dataset is currently private/internal and will be released publicly with the RoMo paper. Licensed under CC BY-NC 4.0, it is suitable for tasks like text-to-3D and motion generation, and includes tools for decoding features to joint positions.




