Parameter constraints for accurate face alignment in the simultaneous inverse compositional algorithm

Face alignment is an important step in many face analysis applications including face recognition, demographics estimation, and expression analysis. Building on the seminal work on Inverse Compositional Image Alignment by Matthews and Baker many improvements have been proposed, although little attention has been paid to constraining parameters during the aligment process. This paper is focused on the SIC optimization technique and compares three methods for parameter regularization. Experimental results carried out on the XM2VTS database show that these modifications improve the rate of convergence and the RMSE, especially in the case of unseen subjects, which is one of the main limitations in SIC-based fitting.