Ambient Lighting Input Images Landmarks L.+Silhouette L.+S.+Colour Ground Truth

We present a method to fit a detailed 3Dmorphable model to multiple images. Our formulation allows the fitting of the model without determining the lighting conditions and albedo of the face, mak-ing the system robust against difficult lighting sit-uations and unmodelled albedo variations such as skin colour, moles, freckles and cast shadows. The cost function employs The model shape prior A small number of landmarks for initialization A monocular silhouette distance cost A stereo colour cost The optimisation consists of multiple runs of a non-linear minimizer. During each run the visibility of all sample points is assumed to stay constant. After some iterations the minimizer is stopped and visi-bility is reevaluated. Model The linear morphable face model was created by registering 200 face scans and performing a PCA on the data matrix to fit a Gaussian probability to the data and reduce the dimensionality of the model. Colour Reprojection Cost The colour reprojection cost measures the image colour dif-ference between the projected positions of sample points in two images. The sample points are spaced out regularly in the projected images