Computing mutual similarity of 3D human faces in nearly linear time.

Abstract

Using three-dimensional scans of human faces has become an emerging technique in studies of human variation, where the quantitative assessment of facial similarity complements the measurement of other somatic traits. While the algorithms for automated registration (geometrical alignment) and similarity measurement of two facial scans are well-known and used in practice, their direct application for batch processing is limited due to computational requirements. The batch N:N analysis, where all pairs of scans in a dataset must be mutually registered and compared, introduces quadratic complexity with computation times reaching hours even for relatively small datasets, making it practically unusable. This paper presents a rapid and accurate approach with nearly linear time complexity. Our solution utilizes properties of facial scan geometry to optimize individual steps. Moreover, the algorithm deals with possible holes and other artifacts in polygonal meshes automatically. Experiments demonstrate that the proposed solution is very fast and sufficiently accurate compared to a precise quadratic-time baseline approach.