Dataset of Structured Queries and Spatial Relations

我们已经收集了大约450，000个相关性注释和53个 结构化查询。在两次传递中，我们都指示注释者采用观察者的参照系。虽然我们的 dataset使用比 [25] 更严格的查询语言，它 由于人类空间概念的使用，仍然具有挑战性 现实世界图像中对象外观的关系和高度变化。虽然，理想情况下，我们也会注释所有 在每张图像中的空间关系，这个过程被证明是 太贵了，因为它按二次比例扩大了wrt。号码 场景中每个关系的对象。因此，我们决定 在一种更具可扩展性的方法上，其中只描述 给出了关系。We have collected approximately 450,000 relevance annotations and 53 structured queries. In two annotation passes, we instructed annotators to adopt the observer's frame of reference. Although our dataset uses a more stringent query language than [25], it remains challenging due to the employment of human spatial concepts related to relational and scale variations of object appearances in real-world images. While ideally we would also annotate all spatial relationships present in every image, this process proved to be overly costly, as it scales quadratically with respect to the number of objects associated with each relationship in the scene. Therefore, we opted for a more scalable approach, where only relational descriptions are provided.

搜集汇总

背景与挑战

背景概述

该数据集包含约45万个相关性注释和53个结构化查询，专注于图像检索和标注任务中的空间关系建模。它采用观察者参照系，以应对现实世界图像中对象外观的挑战，并由马克斯普朗克信息学研究所于2015年发布。