R4R
收藏资源简介:
Room-for-Room(R4R)数据集是由Google Research创建的,旨在扩展Room-to-Room(R2R)数据集,以解决视觉与语言导航(VLN)任务中路径多样性和指令遵循性不足的问题。R4R通过将R2R中的路径对进行拼接,生成更长、更曲折的路径,从而增加路径的复杂性和多样性。该数据集包含233,613条训练路径,平均路径长度为20.6米,显著高于R2R的9.91米。数据来源为R2R中的路径组合,通过算法生成,无需额外的人工标注。R4R的创建过程基于R2R的路径结构,通过连接起点接近的路径对,形成新的复合路径。该数据集主要用于视觉与语言导航任务,旨在提升模型对语言指令的遵循能力和路径规划能力,特别是在复杂路径中的导航性能。
Room-for-Room (R4R) dataset was created by Google Research to expand the Room-to-Room (R2R) dataset, addressing the issues of insufficient path diversity and inadequate instruction-following performance in the Vision-and-Language Navigation (VLN) task. R4R generates longer and more winding navigation paths by concatenating path pairs from the R2R dataset, thereby increasing the complexity and diversity of the paths. This dataset contains 233,613 training paths, with an average path length of 20.6 meters, which is significantly higher than the 9.91 meters of the original R2R dataset. The data is generated via algorithmic combination of paths from the R2R dataset, without requiring any additional manual annotations. The development of R4R is based on the path structure of R2R, connecting path pairs with close starting points to form new composite navigation paths. This dataset is primarily used for VLN tasks, with the goal of enhancing models' ability to follow language instructions and perform path planning, especially their navigation performance in complex path environments.




