User’s Guide

Setting up the Dataset

Download the REFUGE dataset, take note of it’s root directory and configure the data access using the bob command-line configuration utility. For example:

$ conda activate your-bob-env
(your-bob-env) $ bob config set bob.db.refuge.datadir /path/to/root/of/refuge
(your-bob-env) $ bob config show #to check

You can than check if your local version of the dataset is compatible with this interface and has the standard directory tree:

$ conda activate your-bob-env
(your-bob-env) $ refuge checkfiles
checkfiles completed sucessfully


A total of 1200 color fundus photographs are available. All fundus images are stored as JPEG files. The dataset is split 1:1:1 into 3 subsets equally for training, offline validation and onsite test, stratified to have equal glaucoma presence percentage. Training set with a total of 400 color fundus image will be provided together with the corresponding glaucoma status and the unified manual pixel-wise annotations (a.k.a. ground truth). Testing consists of 800 color fundus images and is further split into 400 off-site validation set images and 400 on-site test set images. The two default protocols are:

  1. “default_od” for optic disc

  2. “default_cup” for the optic cup

The images are acquired with two different fundus cameras:

  • Zeiss Visucam 500 (2124x2056 pixels) - For Training

  • Canon CR-2 (1634x1634 pixels) - For Validation and Test


Train and Test images have different resolutions!