RPM: Random Points Matching for Pair wise Face-Similarity

Matching face image pairs based on global features or local analysis on some points found using a key point or fiducial point detector becomes prohibitively difficult in realistic images when there are large pose, lighting, expressions and imaging differences. We develop a new approach that automatically and reliably finds well-matched and useful corresponding points, referred to as homologous points, from randomly initialized points on the two probe images under unrestricted image variations. The procedure obviates the need of using key or fiducial point detector and the over restrictive requirement of image alignment. We then propose a new pair-wise similarity metric that combines the strength of the useful parameters found during the random point matching and the similarity computed using a local descriptor around the homologous points. Our results in a face verification setting on two challenging datasets (‘Labelled Faces in the Wild’ and FacePix) under large pose, expression and imaging variations, show improved performance over the state-of-the-art methods for pair-wise similarity.