HO-3D 手物交互 3D 图像数据集

HO 全称 Hand and Object Poses，这是一个带有标注的手物交互 3D 图像数据集。该数据集包含 66,034 张训练图像（来自 55 个序列）和 11,524 张测试图像（来自 13 个序列）。序列由多机位相机和单机位相机采集，包含来自 YCB 数据集的 10 个对象和 10 个物体。标注通过优化算法自动生成。测试集的手势标注被保留。测试集中算法的准确性可以通过使用 CodaLab 挑战（参见项目页面）的标准度量来评估。测试集和训练集的对象姿态标注公开可用。

HO stands for Hand and Object Poses, which is an annotated 3D image dataset for hand-object interaction. This dataset comprises 66,034 training images (from 55 sequences) and 11,524 test images (from 13 sequences). The sequences are captured by both multi-view and single-view cameras, and include 10 objects from the YCB Dataset as well as 10 objects. All annotations are automatically generated via optimization algorithms. The gesture annotations of the test set are retained. The accuracy of algorithms on the test set can be evaluated using the standard metrics defined in the CodaLab Challenge (refer to the project page for details). The object pose annotations for both the training and test sets are publicly available.

搜集汇总

背景与挑战

背景概述

HO-3D是一个包含77,558张图像（66,034张训练图像和11,524张测试图像）的手物交互3D数据集，主要用于开发基于RGB图像预测遮挡物体时手部姿势的算法。数据集包含多机位和单机位采集的序列，标注通过优化算法自动生成，测试集的手势标注保留以供评估使用。