Expression Invariant Face Recognition with a 3DMM
Brian Amberg · 2008
We introduce a method for expression invariant face recognition. A generative 3D Morphable Model (3DMM) is used to separate identity and expression components. The expression re-moval results in greatly increased recognition performance, even on difficult datasets, with-out a decrease in performance on expression-less datasets. It is applicable to any kind of input data, and was evaluated here on textureless range scans. Model The Model was learnt from 175 subjects. We used one neutral expression scan per identity and 50 expression scans of a subset of the sub-jects. The identity model is a linear model build from the neutral scans. f = µ+Mnαn. (1) For each of the 50 expression scans, we calcu-lated an expression vector as the difference be-tween the expression scan and the correspond-ing neutral scan of that subject. This data is al-ready mode-centered, if we regard the neutral expression as the natural mode of expression data. From these offset vectors an additional expression matrixMe was calculated, such that the complete linear Model is f = µ+Mnαn +Meαe (2) The assumption here is, that the face and ex-pression space are linearly independent, such that each face is represented by a unique set of coefficients. Fitting A Robust Nonrigid ICP method was used to fit the model to the data. Robustness was achieved by iteratively reweighting the correspondences and using hard compatability test for the closest points. Fitting was initialized by a simple nose detector and proceeded fully automatic. Distance Measure The Mahalanobis angle between the identity co-efficients αn was used for classification.