3D Face Reconstruction from a Single Image Using Machine Learning Methodology

This paper presents a method for 3D face reconstruction from a single image. To reconstruct a 3D-shape from a single image, we need to model the relationship between images and shapes. The relationship defines a function from image brightness variation to local shape variation. Some existing methods for 3D reconstruction assume a series of reflectance models that try to emulate the relationship based on physical theoretics. However, the models of the existing methods are not able to fit actual world enough. Therefore, we propose heuristic approach, which empirically learns the relationship by statistical machine learning methodology. Firstly, we create surface normal estimators by learning the relationship between image patches of brightness and true surface normals. However, these estimators make misestimation in some cases. To correct them, we additionally create surface normal correctors by learning from image patches, estimated normal patches, and true normals. We conducted some experiments for evaluating this method. These results showed efficiency of our approach.