OakInk

OakInk包含OakInk图像和OakInk形状数据集。
OakInk图像包含230K个图像帧，总共捕获12个对象，与32个类别的100个对象进行多达5个面向意图的交互。对象姿势是从MoCap系统捕获的，而MANO手姿势是从2D关键点注释中拟合的。基于现实世界人类演示中的手对象姿势，我们通过一种新颖的方法将现实世界对象上的手姿势转移到具有类似能力的虚拟对象: Tink。所有真实世界和转移的交互都构成了基于几何的数据集: OakInk-Shape。OakInk形状包含62k不同的手对象姿势和模型。

OakInk encompasses two core datasets: OakInk Images and OakInk Shapes. OakInk Images includes 230K image frames that capture a total of 12 real-world objects, and features up to 5 intent-oriented hand-object interactions involving 100 objects across 32 categories. The poses of these real-world objects were captured using a motion capture (MoCap) system, while MANO hand poses were fitted from 2D keypoint annotations. Drawing on hand-object poses from real-world human demonstrations, we transfer hand poses from real-world objects to virtual objects with comparable functionalities via a novel approach named Tink. All real-world and transferred hand-object interactions form the geometry-based OakInk-Shape dataset. OakInk Shapes contains 62K distinct hand-object poses and associated models.