RoMoDataset/RoMo-HML-263
收藏资源简介:
RoMo-HML-263 是 RoMo 身体语料库的 263 维 HumanML3D 运动特征表示,配有丰富的多级文本描述。它是用于训练和评估基于 HumanML3D 特征集模型的直接配套数据集。该版本包含 810,281 个剪辑,这些剪辑是 RoMo 论文分发的官方 HumanML3D-263 向量(new_joint_vecs)。数据集采用 Parquet 格式(zstd 压缩分片),包含训练、验证和测试分割(分别为 689,147、80,729 和 40,405 个序列)。每个序列以 30 fps 的帧率存储,特征维度为 263,并具有 5 个级别的文本描述(从标签到段落)。
RoMo-HML-263 is the RoMo body corpus in the 263-dimensional HumanML3D motion-feature representation, paired with rich multi-level text descriptions. It is the drop-in companion for training and evaluating models built around the HumanML3D feature set. This release contains the 810,281 clips for which the RoMo paper distributed official HumanML3D-263 vectors (new_joint_vecs). The dataset is in Parquet format (zstd-compressed shards) with train, validation, and test splits (689,147, 80,729, and 40,405 sequences respectively). Each sequence is stored at 30 fps, has a feature dimension of 263, and includes 5 levels of text descriptions (from tag to paragraph).




