EDITVAL
收藏资源简介:
EDITVAL是一个标准化的基准数据集,用于评估基于文本指导的图像编辑方法。该数据集由马里兰大学创建,包含648个独特的图像编辑操作,涵盖19个来自MS-COCO的类别,旨在解决图像编辑中的多样性和复杂性问题。数据集的创建过程涉及精心挑选的测试图像和手动定义的可编辑属性,以及两种标准化的评估流程:自动化评估和大规模人类研究。EDITVAL的应用领域包括设计、制造和工程,以及机器学习研究,如模型解释性和模型在分布外鲁棒性的改进。
EDITVAL is a standardized benchmark dataset for evaluating text-guided image editing methods. Developed by the University of Maryland, this dataset contains 648 unique image editing operations spanning 19 categories from MS-COCO, and is designed to address the diversity and complexity challenges in image editing. The creation of the dataset involves carefully selected test images and manually defined editable attributes, as well as two standardized evaluation pipelines: automated evaluation and large-scale human studies. The application scenarios of EDITVAL include design, manufacturing, engineering, and machine learning research such as model interpretability and improving the out-of-distribution robustness of models.




