Gabor-based Feature Point Tracking with Automatically Learned Constraints ?
Jan Wieghardt, Christoph von der Malsburg · 2002
All point tracking mechanisms sometimes fail due to ambiguities in the visual data, a problem which can be alleviated by introducing model knowledge in the form of constraints on groups of feature points. Starting from a point tracking mechanism based on Gabor phases we introduce model constraints, on the one hand by posterior regularization (externally) and on the other hand by incorporating them directly into the tracking mechanism (internally). In the special case of facial feature tracking we show how the necessary model knowledge expressed in the constraints can be learned without explicit user interaction. To this end typical transformations of point groups are learned from noisy but automatically determined correspondences via principal component analysis.