GAPartManip
收藏资源简介:
GAPartManip是一个大规模的以部件为中心的数据集,专门用于材料无关的铰接物体操作。该数据集由北京大学计算机科学学院等机构创建,包含19种常见的家用铰接物体类别,总计918个物体实例,240,000张照片级真实感渲染图像,以及80亿个场景级别的可操作交互姿态。数据集的创建过程结合了物理基础的IR图像渲染和部件导向的可操作交互姿态标注,旨在解决现有方法在深度感知和姿态检测中的不足。GAPartManip的应用领域主要集中在家庭场景中的铰接物体操作,旨在提高深度感知和可操作交互姿态预测的性能,特别是在模拟和真实世界场景中的零样本模拟到真实世界的转移。
GAPartManip is a large-scale part-centric dataset dedicated to material-agnostic articulated object manipulation. Created by institutions including the School of Computer Science at Peking University and other relevant research bodies, it encompasses 19 common household articulated object categories, totaling 918 object instances, 240,000 photorealistic rendered images, and 8 billion scene-level actionable interaction poses. The dataset construction combines physics-based IR image rendering and part-oriented actionable interaction pose annotation, aiming to address the limitations of existing methods in depth perception and pose detection. GAPartManip is primarily targeted at articulated object manipulation tasks in household scenarios, with the goal of improving the performance of depth perception and actionable interaction pose prediction, especially for zero-shot sim-to-real transfer across both simulated and real-world settings.




