On Evaluating Pixel-Level Face Image Quality Assessment
A decisive factor for face recognition performance is face image quality (FIQ). It describes the utility of face images for automatic recognition. While this FIQ has conventionally been considered as a scalar for the whole image, emerging works suggest assessing pixel-level FIQs to provide higher explainability. However, the value of pixel-level qualities as a measure of utility (value for recognition) is not yet investigated. In this work, we address this by presenting two evaluation schemes, deletion evaluation curve (DEC) and insertion evaluation curve (IEC). The DEC investigates the change in recognition performance as pixels are deleted based on their quality. Complementary, the IEC reports the change in recognition performance as pixels are inserted based on their quality into a blurred image. Since pixel-level face image quality assessment (PLFIQA) methods assign high values to pixels that contain discriminant information, the recognition performance should decrease or increase when they are removed or added, respectively. We have successfully demonstrated the proposed evaluation scheme on two face recognition solutions by comparing a recently proposed PLFIQA method to a random baseline. With the growing interest in explainable face recognition, the proposed metrics will enable adequate comparison of future advances in pixel-level quality assessment.