Gaussian Conditional Random Fields for Face Recognition

We propose a Gaussian Conditional Random Field (GCRF) approach to modeling the non-stationary distortions that are introduced from changing facial expressions during acquisition. While previous work employed a Gaussian Markov Random Field (GMRF) to perform deformation tolerant matching of periocular images, we show that the approach is not well-suited for facial images, which can contain significantly larger and more complex deformations across the image. Like the GMRF, the GCRF tries to find the maximum scoring assignment between a match pair in the presence of non-stationary deformations. However, unlike the GMRF, the GCRF directly computes the posterior probability that the observed deformation is consistent with the distortions exhibited in other authentic match pairs. The difference is the inclusion of a derived mapping between an input comparison and output deformation score. We evaluate performance on the CMU Multi-PIE facial dataset across all sessions and expressions, finding that the GCRF is significantly more effective at capturing naturally occurring large deformations than the previous GMRF approach.