Extracting motion from images: Robust optic flow and structure from motion

Abstract

The first part of this paper summarises recent work in developing highly robust estimators and to apply these to computer vision applications - specifically for optic flow, range image fitting and segmentation, and image motion estimation. The second part of the paper considers the so called “direct” methods for structure from motion (through matrix factorization), though the methods and principles are also applicable to other areas (face recognition, customer characterization and recommender systems etc.).