i-CIR
收藏资源简介:
i-CIR是一个实例级组合图像检索数据集,旨在检索在文本查询定义的修改下,包含与视觉查询相同特定对象的图像。该数据集包含了202个对象实例和750K张图像,每个实例有1-46个图像查询和1-5个文本修改,共有1883个组合查询。数据集通过半自动化的方式选择了大量的硬负样本,以保持其挑战性。此外,数据集的构建过程结合了人工输入和自动图像检索,使用了LAION数据集中的图像。i-CIR数据集适用于训练和评估组合图像检索方法,旨在解决现有数据集中存在的问题,如标签模糊、假阴性率高、文本查询不充分等。
i-CIR is an instance-level compositional image retrieval dataset designed to retrieve images that contain the same specific object as the visual query under modifications defined by text queries. This dataset includes 202 object instances and 750K images, with each instance having 1 to 46 image queries and 1 to 5 text modifications, amounting to a total of 1,883 compositional queries. A large number of hard negative samples are selected for the dataset via a semi-automated approach to sustain its challenge level. Furthermore, the dataset's construction process combines manual annotation and automated image retrieval, utilizing images from the LAION dataset. The i-CIR dataset is intended for training and evaluating compositional image retrieval methods, and aims to resolve issues present in existing datasets, such as ambiguous labels, high false negative rates, and insufficient text queries.




