An improved 3D Bilinear Multidimensional Morphable Models used in 3D face recognition

This paper proposes the improved 3D Bilinear Multidimensional Morphable Models (BMMMs) which can be used in face recognition, etc. In each level of the multidimensional model, we introduce the bilinear programming model for the model matching problem as well as a globally optimal algorithm for bilinear programs. Both tight convex relaxations of the objective function and a convergent branching strategy contribute to the success of the globally optimal algorithm for bilinear programs. Experiments with synthetic data validate that the 3D Bilinear Multidimensional Morphable Models outperform the 3D Morphable Models in optimality, model matching speed, convergence rate and robustness to noise and outliers.