Combining Features and Intensity for Wide-Baseline Non-Rigid Surface Registration

Abstract

Non-rigid surface registration consists in estimating the deformation of a known flat surface, usually by fitting a warp such as a Thin-Plate Spline or a Free-Form Deformation. Common techniques are split in two categories: (1) pixel-based surface tracking where important deformations can be estimated over a video sequence as long as the frame to frame steps are small and (2) wide-baseline feature-based registration, able to directly handle large deformations at the cost of reduced accuracy. We introduce a new direct data term robustly merging feature and pixel-based costs in a pyramidal variational approach. By using a robust estimator we achieve an implicit optimal filtering of features and automatic balancing between the two terms.