3D Reconstruction of the Human Face using Multi-stereo Camera Network

This thesis presents a method to reconstruct a human face in 3D from its multiple 2D images acquired using multiple cameras arranged in a network. The image acquisition setup consists of three stereo pairs of cameras and two off-the-shelf projectors. Since a human face lacks many trackable features, a single structured light pattern was projected on the face to artificially create sufficient number of feature points. The image acquisition was instantaneous, whereas the 3D reconstruction — a computation-heavy process, involving minimization of sum of squared reprojection error — was performed offline. Minimization of the underlying objective function, known to possess large number of local minima, was carried out using genetic algorithm. Furthermore, The cost function of the camera network was defined in a way that allows us to get the 3D information over a large part of the periphery of the face. Our method achieved true-to-scale reconstruction with satisfactory accuracy, as revealed by measurements of certain key facial features. Furthermore, inspection of the reconstructed point cloud from different perspectives provided visual confirmation that shape was also preserved