M-VQA: Visual Question Answering under Image Distortions
Dataset Description
M-VQA is the official dataset of the M-VQA Challenge, an ACM Multimedia 2026 Grand Challenge on evaluating visual question answering systems under image distortions.
The dataset contains 12,400 image-question-answer samples:
400 original images;
12,000 distorted images generated from the original images using 30 distortion types;
five distortion severity levels, numbered from 4 to 8.… See the full description on the dataset page: https://huggingface.co/datasets/AIBench/M-VQA.
