3D Face Reconstruction from a Single Frontal Face Image by Robust Cascaded Regression

Abstract

In this paper, we present a new learning based 3D face reconstruction approach based on robust cascaded regression to reconstruct the 3D face from a single 2D frontal face image. The approach represents the regression between 2D and 3D faces with a strong regressor which is comprised of a number of trained weak regressors (ferns) in additive way, with each fern calculating the regression between 2D and 3D face shape increments. Then, the reconstructed 3D face is a combination of a number of outputs of ferns, enabling us to get a nonparametric 3D face model. Given a 2D face image and an initial 3D face, the approach reconstructs the 3D face by updating the initial 3D face from coarse to fine. Moreover, a new shape incremental feature is adopted to form good ferns. The advantages of the proposed approach are: 1) it represents the face with a non-parametric model, and updates the 3D face with a cascaded structure, 2) instead of fixed feature, it introduces a novel shape incremental feature which can explore the global and local information from the input 2D face with the current estimated 3D face in each stage. We conduct the experiments on BJUT_3D face database to verify the proposed approach.