Rel3D
收藏资源简介:
Rel3D是由普林斯顿大学创建的首个大规模人类标注的3D空间关系数据集。该数据集包含9990个3D场景,每个场景中包含两个物体,这些物体要么满足一个空间关系(主体-谓词-对象),要么不满足。数据集的内容包括深度、分割掩码、物体位置、姿态和尺度等丰富的几何和语义信息。创建过程涉及众包工作者在Amazon Mechanical Turk上根据指导操作物体,并由独立工作者验证空间关系是否成立。Rel3D的应用领域包括机器人导航、物体操作和人类机器人交互,旨在解决3D空间关系理解和预测的问题。
Rel3D is the first large-scale human-annotated 3D spatial relation dataset created by Princeton University. This dataset contains 9990 3D scenes, each with two objects that either satisfy a spatial relation (subject-predicate-object) or do not. It includes rich geometric and semantic information such as depth maps, segmentation masks, object positions, poses and scales. The dataset's creation involved crowdworkers manipulating objects on Amazon Mechanical Turk following provided guidelines, with independent workers verifying whether the spatial relations hold. Rel3D has application domains including robot navigation, object manipulation and human-robot interaction, and aims to address the problem of 3D spatial relation understanding and prediction.

- 1Rel3D: A Minimally Contrastive Benchmark for Grounding Spatial Relations in 3D普林斯顿大学 · 2020年



