Pascal-EA

Pascal-EA是由北京邮电大学构建的基准数据集，用于评估语义分割模型在不同对象和图像属性变化下的鲁棒性。该数据集通过精确控制结构信息的掩码保留属性编辑管道来编辑真实图像的视觉属性，从而使原始分割标签可用于编辑后的图像。Pascal-EA涵盖了从传统闭集模型到最近的开放词汇大型模型的广泛语义分割模型，旨在揭示模型对局部和全局属性变化的敏感性，并强调对象属性在提高分割鲁棒性中的重要性。

Pascal-EA is a benchmark dataset constructed by Beijing University of Posts and Telecommunications to evaluate the robustness of semantic segmentation models against variations in different objects and image attributes. This dataset edits the visual attributes of real images through a masked attribute editing pipeline that precisely controls the retention of structural information, thereby allowing the original segmentation labels to be applied directly to the edited images. Pascal-EA encompasses a broad spectrum of semantic segmentation models, ranging from traditional closed-set models to recent open-vocabulary large-scale models, with the goal of uncovering the sensitivity of these models to local and global attribute variations and emphasizing the significance of object attributes in improving segmentation robustness.