Optimal Multi-frame Correspondence with Assignment Tensors

Establishing correspondence between features of a set of images has been a long-standing issue amongst the computer vision community. We propose a method that solves the multi-frame correspondence problem by imposing a rank constraint on the observed scene, i.e. rigidity is assumed. Since our algorithm is based solely on a geometrical (global) criterion, it does not suffer from issues usually associated to local methods, such as the aperture problem. We model feature matching by introducing the assignment tensor, which allows simultaneous feature alignment for all images, thus providing a coherent solution to the calibrated multi-frame correspondence problem in a single step of linear complexity. Also, an iterative method is presented that is able to cope with the non-calibrated case. Moreover, our method is able to seamlessly reject a large number of outliers in every image, thus also handling occlusion in an integrated manner.