A proposal of approaches using deep neural networks and geometric methods for 3D face reconstruction from a single image
Oussema Bouafif · theses.fr (ABES) · 2021
3D face reconstruction from a 2D image is a fundamental problem in computer vision that is attracting considerable interest owing to its various potential applications such as surveillance, health, video games, cinema, etc. This thesis presents two hybrid approaches for 3D face reconstruction from a 2D color image that combine deep learning and geometric techniques. To deal with the lack of data needed to train the neural networks, a 3D synthetic human heads generator is designed. It allowed us to provide a facial image database with several maps that contain facial geometry characteristics for each example. Both 3D face reconstruction approaches use CNN to produce two maps from a human face image. The first approach generates a pixel-wise normal map and an image of the magnitude of the depth gradient. Subsequently, using these maps, the 3D facial geometry is recovered by applying a normals integration process based on weighted least squares method. In the second approach, the neural network produces a landmarks map and a pixel-wise normal map similar to the one produced in the first method. Landmarks are used for pose computation and the initialization of an optimization problem, which in turn, reconstructs the 3D head geometry by using a 3D morphable model (3DMM) and normal vector fields.