Automatic impression transformation of faces in 3D shape - a perceptual comparison with processing on 2D images

This paper describes an attempt to transform a 3D model of a person's face to produce an intended change of impression. 3D shape and surface texture of faces are represented by high-dimensional vectors automatically extracted from the 3D data captured by a range finder, and variations among a set of faces are coded by applying principal component analysis. The relationship between the coded representation and the attribute of faces along a given impression dimension is analyzed to obtain an impression transfer vector. Here, we propose a method using this impression transfer vector to manipulate 3D faces in order to transform impressions. Experimental results on transformation of gender impressions confirmed the superiority of manipulating the 3D information of faces over a previous approach using only 2D face information