GeneCIS

GeneCIS是一个用于评估模型适应不同图像相似性条件的基准数据集。该数据集通过重新利用现有的公共数据集构建，包含四个评估任务，旨在测试模型在零射击评估中的表现。GeneCIS涵盖了广泛的实际用例，并设计用于评估模型对开放集相似性条件的适应能力。数据集的构建考虑了用户可能对场景中的对象或属性感兴趣的情况，以及条件可能关注图像的特定方面或指定图像变化的情况。GeneCIS的目的是推动对条件图像相似性问题的研究，特别是在模型如何灵活适应不同相似性概念方面。GeneCIS is a benchmark dataset for evaluating model adaptation to varying image similarity conditions. Constructed by repurposing existing public datasets, it comprises four evaluation tasks intended to test model performance in zero-shot evaluations. GeneCIS covers a broad spectrum of real-world use cases and is designed to assess a model's capacity to adapt to open-set similarity conditions. The dataset was developed with consideration for scenarios where users may focus on objects or attributes within a scene, as well as cases where conditions center on specific aspects of an image or specify image modifications. The primary objective of GeneCIS is to promote research on conditional image similarity issues, especially with respect to how models can flexibly adapt to different similarity notions.