anonymous-submission-usage/RoomTour3D

RoomTour3D是一个用于训练开放世界导航代理的视频指令数据集。该数据集包含1847个房间游览场景的几何感知视频指令数据，并提供了3D场景重建、相对深度估计、开放世界对象标签和定位等中间产品。数据集的注释格式包括COLMAP重建、RAM对象标签、相对深度估计和轨迹信息。数据集还提供了视频下载链接、退出机制和许可证信息。

This is a video-instruction dataset for anonymous review, aimed at training open-world navigation agents. The dataset contains geometry-aware video-instruction data from 1847 room tour scenes, along with intermediate products such as 3D scene reconstruction using COLMAP, relative depth estimation, open-world object tags, and localization. The annotation format of the dataset details the structure of different types of data files, including colmap_reconstruction, ram_grounding_p1, relative_depth_estimation, and trajectories.