Computer Science · Computer Vision and Pattern Recognition
3D Face Reconstruction with A Four Camera Acquisition
Davide Onofrio, Stefano Tubaro, Antonio Rama, Francesc Tarres
See on the mapAbstract
In this paper we present a method for 3D face reconstruction based on the use of a four camera acquisition system. Our method determines correspondences between surface patches on different views through a modeling of depth maps based on Markov Random Fields (MRFs). In order to reduce the occurrence of outliers the MRF-based modeling is bound to satisfy the epipolar constraint. We apply the belief propagation algorithm to the MRF model in order to perform a maximum-a-posteriori estimation of such correspondences. The use of a four camera system improves the accuracy of the obtained 3D model, in fact the occlusions are remarkably reduced, and moreover the extension of the reconstructed face is wider than in more traditional two or three camera systems, this is particularly useful in some 3D-based face recognition systems that work with only a limited regions of a face.