RoomTour3D
收藏资源简介:
RoomTour3D是一个基于网络房间游览视频的几何感知视频指令数据集,旨在解决视觉与语言导航(VLN)任务中训练数据多样性和规模有限的问题。该数据集包含约10万条开放式描述丰富的轨迹和约20万条指令,涵盖1847个真实房间环境。数据集通过3D重建技术获取行走路径的几何信息,并结合房间类型、物体位置和场景3D形状等额外信息。创建过程中,利用自动化的视频处理管道生成几何感知的行走轨迹和空间上下文文本指令。RoomTour3D主要应用于视觉与语言导航任务,旨在提升导航代理在开放世界环境中的表现,解决复杂室内环境中的导航问题。
RoomTour3D is a geometry-aware video instruction dataset built on web-based room tour videos, aiming to address the limitations in diversity and scale of training data for Vision-and-Language Navigation (VLN) tasks. This dataset contains approximately 100,000 trajectories with rich open-ended descriptions and around 200,000 instructions, covering 1,847 real-world room environments. It acquires geometric information of walking paths via 3D reconstruction techniques, and incorporates additional metadata including room types, object positions and 3D scene shapes. During its development, an automated video processing pipeline is utilized to generate geometry-aware walking trajectories and spatially contextual textual instructions. RoomTour3D is primarily applied to Vision-and-Language Navigation tasks, with the goal of enhancing the performance of navigation agents in open-world environments and solving navigation challenges in complex indoor settings.
RoomTour3D 数据集概述
数据集名称
RoomTour3D
数据集描述
RoomTour3D 是一个用于具身导航的几何感知视频指令调优数据集。该数据集专注于视觉与语言导航任务,旨在通过视频指令和几何信息来训练和评估具身导航模型。
主要研究领域
- 视觉与语言导航
- 具身导航
- 多模态学习
- 视频指令调优
数据集作者
- Mingfei Han
- Liang Ma
- Kamila Zhumakhanova
- Ekaterina Radionova
- Jingyi Zhang
- Xiaojun Chang
- Xiaodan Liang
- Ivan Laptev
作者单位
-
- 计算机视觉系,MBZUAI
-
- 中山大学深圳校区
-
- ReLER Lab, AAII, UTS
-
- 中国科学技术大学
相关链接

- 1RoomTour3D: Geometry-Aware Video-Instruction Tuning for Embodied Navigation计算机视觉系,MBZUAI,中山大学深圳校区,AAII,UTS,中国科学技术大学 · 2024年



