Common Objects Out-of-Context (COOCO)
收藏资源简介:
COOCO数据集是一个用于评估多模态模型在场景上下文中整合对象级和场景级视觉信息能力的全新数据集。该数据集包含18,395张图像,分为原始图像、低、中、高相关性和同类型目标条件。数据集通过在COCO-Search18数据集的基础上引入与场景类型具有不同程度语义相关性的目标对象来构建,旨在研究视觉语言模型在场景语义违规情况下的行为。该数据集可用于评估模型在不同场景上下文和视觉噪声条件下对目标对象识别的依赖程度,以及模型在生成引用表达时的鲁棒性。
The COOCO Dataset is a novel dataset developed to evaluate the capability of multimodal models to integrate object-level and scene-level visual information within contextual scenarios. The dataset comprises 18,395 images, categorized into four conditions: original images, conditions with low, medium, and high semantic correlation targets, and same-type target conditions. Built upon the COCO-Search18 dataset by introducing target objects with varying degrees of semantic relevance to the scene type, this dataset aims to study the behaviors of vision-language models under scenarios of scene semantic violations. This dataset can be used to assess the extent to which models rely on target object recognition under different scene contexts and visual noise conditions, as well as the robustness of models when generating referring expressions.
数据集概述
基本信息
- 数据集名称:scenereg
说明
- 无其他相关信息。

- 1COOCO -- Common Objects Out-of-Context -- Semantic Violation in Scenes: Investigating Multimodal Context in Referential Communication乌德勒支大学, 特伦托大学 · 2025年



