Human-Preference FashionIQ (HP-FashionIQ)
收藏资源简介:
HP-FashionIQ数据集是针对复合图像检索(CIR)任务创建的一个新数据集,它通过捕获用户在目标图像检索之外的其他图像上的偏好,来评估CIR模型的性能。该数据集包含3,050个查询,每个查询都有两个图像集合,分别由不同的CIR模型检索得到,并由人类评估员进行偏好标注。HP-FashionIQ旨在解决现有CIR数据集的局限性,即仅标注单个目标图像,忽略了其他相关图像的重要性。通过引入人类偏好标注,HP-FashionIQ为评估CIR模型的性能提供了一个更全面和人性化的标准。
The HP-FashionIQ dataset is a novel dataset created for the composite image retrieval (CIR) task, which evaluates the performance of CIR models by capturing user preferences over images other than the target retrieval image. This dataset contains 3,050 queries, each paired with two image sets retrieved by different CIR models, with preference annotations provided by human evaluators. HP-FashionIQ aims to address the limitations of existing CIR datasets, which only annotate single target images and overlook the importance of other relevant images. By introducing human preference annotations, HP-FashionIQ offers a more comprehensive and human-centric standard for evaluating the performance of CIR models.

- 1QuRe: Query-Relevant Retrieval through Hard Negative Sampling in Composed Image RetrievalKAIST · 2025年



