Functional difference predictors (FDPs): measuring meaningful image differences

In this paper we introduce functional difference predictors (FDPs), a new class of perceptually-based image difference metrics that predict how image errors affect the ability to perform visual tasks using the images. To define the properties of FDPs, we conduct a psychophysical experiment that focuses on two visual tasks: spatial layout and material estimation. In the experiment we introduce errors in the positions and contrasts of objects reflected in glossy surfaces and ask subjects to make layout and material judgments. The results indicate that layout estimation depends only on positional errors in the reflections and material estimation depends only on contrast errors. These results suggest that in many task contexts, large visible image errors may be tolerated without loss in task performance, and that FDPs may be better predictors of the relationship between errors and performance than current visible difference predictors (VDPs).