Assessing the match performance of non-ideal operational facial images using 3D image data.

Biometric attributes are unique characteristics specific to an individual, which can be used in automated identification schemes. There have been considerable advancements in the field of face recognition recently, but challenges still exist. One of these challenges is pose-variation, specifically, roll, pitch, and yaw variations away from a frontal image. The goal of this problem report is to assess the improvement of facial recognition performance obtainable by commercial pose-correction software. This was done using pose-corrected images obtained in two ways: 1) non-frontal images generated and corrected using 3D facial scans (pseudo-pose-correction) and 2) the same non-frontal images corrected using FaceVACs DBScan. Two matchers were used to evaluate matching performance namely Cognitec FaceVACs and MegaMatcher 5.0 SDK. A set of matching experiments were conducted using frontal, non-frontal and pose-corrected images to assess the improvement in matching performance, including: 1. Frontal (probe) to Frontal (gallery) images, to generate the baseline 2. Non-ideal pose-varying (probe) to frontal (gallery) 3. Pseudo-pose-corrected (probe) to frontal (gallery) 4. Auto-pose-corrected (probe) to frontal (gallery). Cumulative match characteristics curves (CMC) are used to evaluate the performance of the match scores generated. These matching results have shown better performance in case of pseudo-pose-corrected images compared to the non-frontal images, where the rank accuracy is 100% for the angles which were not detected by the matchers in the non-frontal case. Of the two commercial matchers, Cognitec, which is software optimized for non-frontal models, has shown a better performance in detection of face with angular rotations. MegaMatcher, which is not a pose-correction matcher, was unable to detect greater angles of rotation which are 50° and 60° in pitch, greater than 40° for yaw and for coupled pitch/yaw it was unable to detect 4 out of 8 combinations. The requirements of the facial recognition application will influence the decision to implement pose correction tools