Comparing landmarking methods for face recognition

Good registration (alignment to a reference) is essential for accurate face recognition. We use the locations of facial features (eyes, nose, mouth, etc) as landmarks for registration. Two landmarking methods are explored and compared: (1) the Most Likely-Landmark Locator (MLLL), based on maximizing the likelihood ratio [1], and (2) ViolaJones detection [2]. Further, a landmark-correction method based on projection into a subspace is introduced. Both landmarking methods have been trained on the landmarked images in the BioID database [3]. The MLLL has been trained for locating 17 landmarks and the ViolaJones method for 5 landmarks. The localization error and effects on the equal-error rate (EER) have been measured. In these experiments groundtruth data has been used as a reference. The results are described as follows: 1. The localization errors obtained on the FRGC database are 4.2, 8.6 and 4.6 pixels for the Viola-Jones, the MLLL, and the MLLL after landmark correction, respectively. The inter-eye distance of the reference face is 100 pixels. The MLLL with landmark correction scores best in the verification experiment. 2. Using more landmarks decreases the average localization error and the EER.