IE-Bench
收藏资源简介:
IE-Bench是由北京大学和鹏城实验室联合创建的一个专门用于评估文本驱动图像编辑的数据集。该数据集包含3010条数据,涵盖了多样化的源图像、编辑提示和不同编辑方法的结果,并提供了25名人类受试者的平均意见分数(MOS)。数据集的创建过程包括源图像收集、编辑提示设计、多种编辑方法的应用以及主观评分实验。IE-Bench旨在解决文本驱动图像编辑任务中评估指标与人类感知不一致的问题,为图像编辑领域的研究提供了重要的基准和评估工具。
IE-Bench is a specialized dataset jointly created by Peking University and Peng Cheng Laboratory for evaluating text-driven image editing tasks. This dataset contains 3010 samples, covering diverse source images, editing prompts, and results produced by different editing methods, and provides the Mean Opinion Scores (MOS) averaged from 25 human subjects. The construction process of IE-Bench includes source image collection, editing prompt design, application of multiple editing methods, and subjective scoring experiments. IE-Bench aims to address the inconsistency between existing evaluation metrics and human perception in text-driven image editing tasks, providing an important benchmark and evaluation tool for research in the field of image editing.




