RAID-Dataset: human responses to affine image distortions and Gaussian noise

Image quality datasets are used to train and evaluate predictive models of subjective human perception. However, most existing datasets focus on distortions commonly found in digital media and not in natural conditions. Affine transformations are particularly relevant for study, as they are among the most commonly encountered by human observers in everyday life. This Data Descriptor presents a set of human responses to suprathreshold affine image transformations (rotation, translation, scaling) and Gaussian noise as convenient reference to compare with existing image quality datasets. The responses were measured using well-established psychophysics: the Maximum Likelihood Difference Scaling (MLDS). The set contains responses to 864 distorted images. The experiments involved 210 observers and over 40,000 image quadruple comparisons. The dataset is validated by two facts: (a) the responses reproduce classical absolute detection thresholds of the affine and Gaussian distortions, and (b) the responses to Gaussian distortion are correlated to the Mean Opinion Score (MOS) of conventional image quality databases for that distortion. Moreover, the classical Piéron's law applies to the reaction times of the dataset, and Group-MAD adversarial stimuli reveal that MLDS perceptual scales are more accurate than the conventional MOS.