How effective are landmarks and their geometry for face recognition?

This paper evaluates how biologically meaningful landmarks and their geometry extracted from face images can be used for face recognition. The traditional Procrustes distance is studied for the landmark-based face model. By using complex principal component analysis, we propose a refined Procrustes distance that incorporates statistical correlation of landmarks. Motivated by research results from art, anthropology, and aesthetic surgery where ratios play a significant role in human face descriptions, we also investigate how well-normalized Euclidean distances (special ratios) can be exploited for face recognition. Exploiting symmetry and using principal component analysis, we reduce the number of ratios to 20. In this investigation, we analyze the effectiveness of three well-known similarity measures including the typical l1 norm, l2 norm, and Mahalanobis distance. We also define a new similarity measure called the eigenvalue-weighted cosine (EWC) distance. We evaluate our approach using two standard face databases: the Purdue AR and the FERET database. The former has well-controlled face images and we use it to test the reliability of landmarks. The FERET database is used for performance evaluation of face recognition algorithms. This papers use a performance measure called the cumulative match score, which indicates the percentage of faces that can be eliminated from consideration. Experimental results show that our EWC distance outperforms the l1 and l2 norms and even the Mahalanobis distance. Our study also finds that landmarks and their geometry-based approach can account for variations of face expression and aging very well. Thus, they can be used either in stand-alone mode or in conjunction with other approaches to reduce the search space a priori. 2005 Elsevier Inc. All rights reserved.