OD\/OC segmentation dataset

The Retinal Fundus Glaucoma Challenge (REFUGE) dataset is our primary benchmark and is known for distribution differences between its subsets. We follow the official split with 1\\,200 training, 400 validation and 400 test images, each with expert OD\/OC masks and image-level glaucoma labels.The ORIGA dataset, collected by the Singapore Eye Research Institute, contains 650 images (168 glaucomatous) with OD\/OC annotations and glaucoma diagnosis. We use 260\/65\/325 images for training\/validation\/testing, preserving the original class distribution.RIM-ONE DL is a curated version of the RIM-ONE database, comprising 485 images (313 normal, 172 glaucomatous) with expert-validated OD\/OC masks. We split the data into 193 training, 49 validation and 243 test images to evaluate cross-dataset robustness.