3D Twins Expression Challenge

We describe the 3D Twins Expression Challenge (“3D TEC”) problem in the area of 3D face recognition. The supporting dataset contains 3D scans of pairs of identical twins taken with two different facial expressions, neutral and smiling. The dataset is smaller than the FRGC v2 [1] dataset by approximately a factor of ten, but is still more challenging than the FRGC v2 dataset due to it containing twins with different expressions. This challenge problem will help to push the frontiers of 3D face recognition. Three dimensional face recognition is an active research topic in biometrics [2, 3]. While 2D pictures can be captured quickly, non-intrusively, and easily by widely available cameras, the images are easily affected by variations in pose, lighting conditions, and facial expressions. 3D face recognition is less affected by pose, facial color variations, and illumination changes since the curvature and shape information of the face is available. Since 3D images offer additional information about the face, 3D images combined with 2D images should give better performance than either alone [4]. The best performance in 3D face recognition algorithms has become high enough in large datasets like FRGC v2 that it is difficult to achieve further significant increases in recognition performance. Two problems generally considered to be difficult are variations in expressions and distinguishing between faces of identical twins. We introduce the 3D TEC dataset which consists of 3D scans of 107 pairs of twins taken in a single session, with each subject having a scan of a neutral expression and a smiling expression. The combination of factors related to the facial similarity of identical twins and the variation in facial expression makes this a challenging dataset. Recently, there have been some twins face recognition studies in biometrics research. Phillips et al. [5] assessed the performance of three of the top algorithms submitted to the Multiple Biometric Evaluation (MBE) 2010 Still Face Track [6] on a dataset of twins acquired at Twins Days [7] in 2009 and 2010. They examined the performance using images acquired in the same day, and also images acquired a year apart (i.e., where the face images acquired in the first year are gallery images and the face images acquired in the second year are probe images). They also examined the performance with varying illumination conditions and expressions. They found that results ranged from around 2.5% Equal Error Rate (EER) for images taken in the same day with controlled lighting and neutral expressions, to around 21% EER for gallery and probe images acquired in different years and in different lighting conditions. Sun et al. [8] conducted a study on multiple biometric traits of twins. They found no significant difference in performance using non-twins compared to using twins for their iris biometric system. For their fingerprint biometric system, they observed that performance using non-twins was slightly better than using twins. Additionally, using their face biometric system, they could distinguish non-twins much better than twins.