HO-3D

具有手和对象的 3D 姿势注释的手对象交互数据集。该数据集包含来自总共 68 个序列的 66,034 个训练图像和 11,524 个测试图像。这些序列在多摄像头和单摄像头设置中捕获，包含 10 个不同的主体，从 YCB 数据集中操纵 10 个不同的对象。注释是使用优化算法自动获得的。测试集的手势注释被保留，测试集上算法的准确性可以使用 CodaLab 挑战提交（参见项目页面）通过标准指标进行评估。测试集和训练集的对象姿态注释与数据集一起提供。A hand-object interaction dataset with 3D pose annotations for hands and objects. This dataset contains 66,034 training images and 11,524 test images from a total of 68 sequences. These sequences are captured in both multi-camera and single-camera settings, involving 10 distinct subjects manipulating 10 different objects from the YCB dataset. The annotations are automatically obtained using optimization algorithms. The gesture annotations of the test set are withheld, and the accuracy of algorithms on the test set can be evaluated via standard metrics by submitting to the CodaLab challenge (refer to the project page). Object pose annotations for both the training and test sets are provided along with the dataset.

搜集汇总

背景与挑战

背景概述

HO-3D是一个专注于手对象交互的3D姿势注释数据集，包含来自68个序列的66,034个训练图像和11,524个测试图像，涵盖10个不同主体和10个不同对象，注释通过优化算法自动生成。该数据集适用于计算机视觉任务如目标检测和关键点检测，测试集注释保留用于算法评估，由格拉茨理工大学和古塔夫埃菲尔大学于2020年发布。