3D face alignment in the wild: A landmark-free, nose-based approach

Abstract

We present a methodology for 3D face alignment in the wild, such that only the nose is required as input for assessing the position of the landmarks. Our approach works by first detecting the nose region, which is used for estimating the head pose. After that, a generic face landmark model, obtained by averaging all training images, is rotated, translated and scaled based on the size and localization of the nose. Because little information is needed and there are no refinement steps, our method is able to find suitable landmarks even in challenging poses. While not taking into account facial expressions and specific facial traits, our algorithm achieved competitive scores on the 3D Face Alignment in the Wild (3DFAW) challenge. The obtained results have the potential to be used as rough estimation of the position of the 3D face landmarks in the wild images, which can be further refined by specially designed algorithms.