HPIR

HPIR是一个专注于评估图像检索系统与人类审美偏好对齐的新型数据集。该数据集由东南大学创建，包含150个伪查询，这些查询通过大型语言模型生成，以确保与用户主题分布一致。每个查询结合多个搜索引擎的结果，生成两组各10张图片，供人类标注者评估哪一组更精确和视觉上更吸引人。每组评估重复30次以确保结果的稳健性。HPIR旨在通过人类标注的偏好数据，评估和提升图像检索系统在审美层面的表现，解决现有检索系统在审美和责任AI原则方面的不足。

HPIR is a novel dataset dedicated to evaluating the alignment between image retrieval systems and human aesthetic preferences. Developed by Southeast University, this dataset comprises 150 pseudo-queries generated via large language models, ensuring consistency with the distribution of user topics. For each query, results aggregated from multiple search engines are used to generate two separate sets, each containing 10 images. Human annotators are then tasked with evaluating which set is more accurate and visually appealing. To ensure the robustness of the results, each evaluation round is repeated 30 times. The core goal of HPIR is to assess and improve the aesthetic performance of image retrieval systems, addressing the shortcomings of existing retrieval systems in terms of both aesthetic judgment and responsible AI principles.