Using genetic algorithms in computer vision : registering images to 3D surface model

Abstract

This paper shows a successful application of genetic algorithms in computer vision. We aim at building photorealistic 3D models of real-world objects by adding textural information to the geometry. In this paper we focus on the 2D-3D registration problem: given a 3D geometric model of an object, and optical images of the same object, we need to find the precise alignment of the 2D images to the 3D model. We generalise the photo-consistency approach of Clarkson et al. who assume calibrated cameras, thus only the pose of the object in the world needs to be estimated. Our method extends this approach to the case of uncalibrated cameras, when both intrinsic and extrinsic camera parameters are unknown. We formulate the problem as an optimisation and use a genetic algorithm to find a solution. We use semi-synthetic data to study the effects of different parameter settings on the registration. Additionally, experimental results on real data are presented to demonstrate the efficiency of the method.