D3D-HOI

该数据集包含有关人类与物体互动过程中3D物体姿态和部分运动的标注视频。为了提高数据质量，已过滤掉严重遮挡和标注噪声的视频，并仅使用捕捉到一半关节运动的帧。该数据集规模为159个视频，涵盖4个物体类别。其任务是对关节运动进行估计，并评估细粒度的视觉对应关系。

This dataset contains annotated videos recording 3D object poses and partial motions during human-object interactions. To enhance data quality, videos with severe occlusions and annotation noise have been filtered out, and only frames that capture half of the joint motions are retained. This dataset comprises 159 videos across 4 distinct object categories. The core tasks of this dataset involve joint motion estimation and fine-grained visual correspondence evaluation.