Visual Relations benchmark

为了进一步定量研究我们技术的关系能力，我们引入了一种新的视觉关系基准，简称VR。该基准包括以下模板的320关系: 建筑物 → 国家，国家 → 资本，食品 → 国家，领导者 → 国家和企业 → 公司。之所以选择这些，是因为它们大致是多对一的，即一个国家有许多建筑物，但建筑物仅与一个国家有关。该基准旨在衡量可视化建模关系的能力以及应用现实世界知识来执行任务的能力。

To further quantitatively investigate the relational capabilities of our proposed technology, we introduce a novel visual relationship benchmark, abbreviated as VR. This benchmark encompasses 320 relationship instances adhering to the following templates: building → country, country → capital, food → country, leader → country, and enterprise → company. These templates were selected due to their predominantly many-to-one relational characteristics: for example, a country hosts numerous buildings, yet each individual building is only associated with one specific country. The benchmark is designed to evaluate the ability to visually model relational structures as well as the capacity to apply real-world knowledge to perform the target task.

搜集汇总

背景与挑战

背景概述

该数据集是一个视觉关系基准，包含320种多对一关系模板，旨在评估视觉关系建模和现实世界知识应用能力。它由特拉维夫大学于2022年发布。