R2R因果推断数据集

RoomtoRoom(R2R)是一个用于真实建筑物中基于视觉的自然语言导航的基准数据集,包括平均句长29词的导航指令及对应的路径信息。基于大规模3D数据集Matterport3D，包括高动态范围彩色图像，深度图，全景天空盒，材质网格，区域布局和物体语义分割。每个真实场景包括一组18张1280×1024分辨率的HDR图像，每张图都带6个自由度的相机位姿估计和同一全景下的一个天空盒、整个场景的纹理多边形网格面。使用Voxel Hashing算法做纹理网格重构算法。

RoomtoRoom (R2R) is a benchmark dataset for vision-and-language navigation in real-world buildings, encompassing navigation instructions with an average length of 29 words and their corresponding path information. It is constructed based on the large-scale 3D dataset Matterport3D, which includes high dynamic range (HDR) color images, depth maps, panoramic skyboxes, textured meshes, regional layouts, and object semantic segmentations. Each real-world scene consists of a set of 18 HDR images at a resolution of 1280×1024, with each image paired with 6-degree-of-freedom (6DoF) camera pose estimates, a skybox corresponding to the same panorama, and the textured polygonal mesh of the entire scene. The Voxel Hashing algorithm is employed for textured mesh reconstruction.

搜集汇总

背景与挑战

背景概述

R2R因果推断数据集是一个用于真实建筑物中基于视觉的自然语言导航的基准数据集，基于大规模3D数据集Matterport3D构建，包含导航指令、路径信息以及高动态范围彩色图像、深度图等多种视觉数据。该数据集旨在支持因果推断研究，适用于视觉导航任务，数据量较小为1.75MB，由4个文件组成。