ORIC
收藏资源简介:
ORIC是一个用于评估大型视觉语言模型在视觉上下文不一致情况下物体识别性能的基准数据集。该数据集由加州大学圣地亚哥分校的研究团队创建,旨在解决大型视觉语言模型在识别与预期背景不符的物体时出现的错误问题,例如物体误识别和幻觉。数据集包含1000个图像,通过LLM引导采样和CLIP引导采样两种策略构建,用于评估模型在识别背景与物体不一致的情况下的性能。该数据集的应用领域包括计算机视觉、自然语言处理和人工智能等领域,旨在帮助研究人员更好地理解视觉语言模型在处理复杂视觉上下文时的局限性和挑战。
ORIC is a benchmark dataset developed to evaluate the object recognition performance of large vision-language models under visual context inconsistency. It was created by a research team at the University of California, San Diego, aiming to address errors such as object misrecognition and hallucinations that large vision-language models make when recognizing objects that do not match their expected backgrounds. The dataset contains 1,000 images and is constructed via two strategies: LLM-guided sampling and CLIP-guided sampling, to evaluate the models' performance in recognizing objects inconsistent with their surrounding contexts. Its application areas cover computer vision, natural language processing, artificial intelligence and other related fields, and it is designed to help researchers better understand the limitations and challenges that vision-language models encounter when processing complex visual contexts.

- 1通过加州大学圣地亚哥分校 · 2025年



