HPS Dataset (Human POSEitioning System Dataset)

HPS 数据集是 3D 人类与大型 3D 场景（300-1000 $m^2$，最高 2500 $m^2$）交互的集合。该数据集包含从头戴式相机捕获的图像，以及参考 3D 姿势和人在预扫描 3D 场景中的位置。拍摄了8个大场景中的7个人进行运动、阅读、吃饭、讲课、使用电脑、煮咖啡、跳舞等活动。该数据集提供超过 300K 的同步 RGB 图像以及参考 3D 姿势和位置。该数据集可用作具有场景约束的以自我为中心的跟踪测试平台，了解人类如何长时间在大场景中交互和移动，以及了解人类如何处理到达他们眼睛的视觉输入。

The HPS dataset is a collection of 3D human interactions with large 3D scenes (ranging from 300 to 1000 $m^2$, with a maximum size of 2500 $m^2$). It contains images captured by head-mounted cameras, along with reference 3D poses and the positions of humans within pre-scanned 3D scenes. Seven individuals across eight large scenes were recorded performing activities such as locomotion, reading, eating, lecturing, using computers, making coffee, dancing, and other actions. The dataset provides over 300K synchronized RGB images, together with reference 3D poses and positions. This dataset can serve as an egocentric tracking testbed with scene constraints, for studying how humans interact and move within large scenes over extended periods, as well as how humans process visual inputs reaching their eyes.

搜集汇总

背景与挑战

背景概述

HPS数据集是一个用于3D人体姿态估计和自定位的数据集，包含从头戴相机捕获的超过30万张同步RGB图像，以及参考3D姿势和人在大型3D场景中的位置信息。该数据集覆盖多个场景和活动，由图宾根大学等机构于2021年发布，适用于研究场景约束下的以自我为中心跟踪和人类行为分析。