HumanML3D (Restore)

该数据集名为HumanML3D（Restore），是原始HumanML3D数据集的改编版本，包含了从AMASS源运动中提取的片段以及相应的描述。数据增强手段包括对运动片段及其对应查询进行左右镜像处理。在规模方面，改编后的片段包含了时间信息，并且为了避免训练集和测试集之间的重叠，进行了重新划分。该数据集的任务是针对人体运动的时间性句子定位。This dataset, named HumanML3D (Restore), is a modified variant of the original HumanML3D dataset. It contains motion clips extracted from the AMASS source motion corpus and their corresponding textual descriptions. Data augmentation strategies include left-right mirroring applied to both the motion clips and their matching textual queries. The modified motion clips incorporate temporal information, and the dataset has been re-partitioned to eliminate overlaps between the training and test splits. The core task of this dataset is temporal sentence localization for human motion.