Single image reconstruction of human faces using database of depth images

Abstract

Despite increasing accessibility of the 3D scanning devices, acquisition of the 2D photographs is still more widely used method to capture human faces. However, many virtual scenarios require 3D representation of the human face. We present an automatic method of the human face reconstruction from a single frontal photograph. Our approach utilizes a database of the precomputed depth images, which we generated from the FIDENTIS Database. This allows us to the avoid lighting constraints, that are often posed on the input data for 2D reconstruction of the human faces. We assume the input image comes annotated with a set of landmarks and we then proceed to reconstruct a 3D representation of the input image by choosing the most appropriate depth image from the database. While the precision of the landmark localization influences the reconstruction, we show, that using the automatic landmark detection methods we can reconstruct a recognizable 3D human face. We further show, that despite a small number of the depth images, the generated meshes are morphologically diverse to a degree, which is suitable for virtual simulations and computer games.