A Bilinear Illumination Model for Robust Face Recognition
Speaker: Baback Moghaddam , Mitsubishi Electric Research Laboratories
We propose a novel technique for computing illumination subspaces for 3D faces from the joint statistics of dense reflectance samples measured from a population of faces under variable lighting. The complex and nonlinear reflectance properties of real human faces is (partially) represented by a bilinear model. We use a higher-order singular value decomposition ("TensorSVD") yielding a compact illumination subspace indexed by face geometry (shape coefficients). The bilinear model is fitted by a re-estimation procedure for minimizing an image-based reconstruction error and leads to the recovery of a shape-specific illumination subspace for a novel individual, obtained from a single 2D photograph. The robustness of our model is evaluated with standard datasets of variable lighting where its performance is shown to be very competitive with past recognition algorithms that rely on analytic illumination models (like spherical harmonics) or learning-based methods requiring multiple training images per subject.