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。

The OakInk dataset encompasses two subsets: OakInk Images and OakInk Shapes. OakInk Images contains 230,000 image frames, which capture a total of 12 objects and up to 5 goal-oriented interactions with 100 objects belonging to 32 categories. The object poses were captured via a MoCap system, while the MANO hand poses were fitted from 2D key-point annotations. Based on the hand-object poses from real-world human demonstrations, we transferred the hand poses on real-world objects to virtual objects with similar capabilities via a novel method named Tink. All real-world and transferred interactions constitute the geometry-based dataset: OakInk-Shape. OakInk Shapes contains 62,000 distinct hand-object poses and models.

搜集汇总

背景与挑战

背景概述

OakInk是一个大规模的手-对象交互理解知识库，包含OakInk图像（230K图像帧）和OakInk形状（62K手对象姿势和模型）两个子数据集，由上海交通大学·上海期智研究院于2022年发布。