3D face reconstruction from images under arbitrary illumination using Support Vector Regression

We present an algorithm for 3D face reconstruction from multiple images under arbitrary illumination. A computer screen is used to illuminate a face from different angles. Three images under different illuminations are used to compute its basis vectors using SVD. The first basis vectors from training faces are projected to a PCA subspace and used as input patterns to train multiple Support Vector Machines. For training, the ground truth 3D face models acquired with a laser scanner are projected to a 13 dimensional PCA subspace and used as output labels. A separate function is learned using Support Vector Regression to estimate each of the 13 parameters of the 3D face. During testing, three images of an unknown face under arbitrary illumination are used to estimate its 3D model. Experiments were performed on 106 subjects and quantitative results are reported by comparing the reconstructed 3D faces to ground truth laser scans. Qualitative results are also reported on the Yale B database.