FunPiQ: A New Benchmark for Pixel-Level Quality Assessment in Fundus Images

FunPiQ is a pixel-level annotation dataset for fundus image quality assessment (FIQA). We sampled images with quality issues from three image-level FIQA datasets: EyeQ, BRSET, and mBRSET. This repository provides only the pixel-level annotations; the corresponding fundus images are available from the original dataset providers. Dataset description The three folders EyeQ, BRSET, and mBRSET contain annotations corresponding to images from each source dataset. Each original fundus image can be identified by matching the annotation filename stem. Label definitions and mask format The masks contain three quality labels plus background: 0: background 1: good 2: usable 3: bad The annotations are stored as 8-bit indexed PNG files. The stored pixel values are the class indices 0–3 and are displayed using a black, green, orange, and red palette. They can be read using Python and Pillow (PIL): from PIL import Image import numpy as np # Load the indexed PNG mask mask_path = "path/to/annotation.png" pil_mask = Image.open(mask_path) # Convert to a NumPy array containing the exact class indices mask_array = np.array(pil_mask) # Optionally retrieve the color palette for visualization palette = pil_mask.getpalette() Original datasets You can download the original datasets from the following sources: EyeQ BRSET mBRSET Please note that EyeQ is itself a reannotation of EyePACS.