Statistical optimization and geometric inference in computer vision
Kenichi Kanatani · Philosophical Transactions of the Royal Society A Mathematical Physical and Engineering Sciences · 1998
This paper gives a mathematical formulation to the computer vision task of inferring three–dimensional structures of the scene based on image data and geometric constraints. Introducing a statistical model of image noise, we define a geometric model as a manifold determined by the constraints and view the problem as model fitting. We then present a general mathematical framework for proving optimality of estimation, deriving optimal schemes, and selecting appropriate models. Finally, we illustrate our theory by applying it to curve fitting and structure from motion.