M-FashionIQ; M-CIRR
收藏资源简介:
M-FashionIQ和M-CIRR是由山东大学、香港理工大学等机构构建的多模态修改数据集,旨在解决组合图像检索(CIR)任务中实体覆盖不足和子句-实体错位问题。数据集基于FashionIQ和CIRR扩展,通过多模态大语言模型(MLLM)生成并人工验证的多修改文本(MMT)增强数据复杂性。M-FashionIQ聚焦时尚领域,涵盖服装形状、颜色等属性;M-CIRR为开放域数据集,强调多对象细粒度描述。数据集通过自动标注与人工审核结合,确保高质量和实用性,适用于复杂多修改场景下的CIR模型训练与评估。
M-FashionIQ and M-CIRR are multimodal modification datasets developed by institutions including Shandong University and The Hong Kong Polytechnic University, aiming to address two critical challenges in the Composed Image Retrieval (CIR) task: insufficient entity coverage and clause-entity misalignment. These datasets are expanded from the original FashionIQ and CIRR datasets, with their complexity enhanced by Multi-Modification Texts (MMT) generated by Multimodal Large Language Models (MLLM) and validated via manual inspection. M-FashionIQ is tailored for the fashion domain, covering attributes such as clothing silhouette and color; whereas M-CIRR is an open-domain dataset that emphasizes fine-grained descriptions of multiple objects. Leveraging a hybrid pipeline combining automatic annotation and manual review, the datasets ensure high quality and practicality, and are suitable for training and evaluating CIR models in complex multi-modification scenarios.




