Supplementing Bundle Adjustment with Evolutionary Algorithms

Bundle Adjustment is a common technique to improve results of any multiple view reconstruction algorithm to obtain 3D structure for computer vision and computer graphics. If the error of a reconstruction can be expressed by an error function, this function can be minimized by numerical methods such as the Levenberg-Marquardt algorithm. By this means, the reconstruction can often be significantly improved. Unfortunately, there is no guarantee for the detected minimum of being a global minimum, since numerical optimization algorithms converge at local minima. The idea presented in this paper is to support the optimization process by evolutionary algorithms. While the existence of fast Levenberg-Marquardt algorithms allow for an obviously faster solution than evolutionary algorithms, the latter can mitigate their disadvantage of getting trapped in local minima. We demonstrate the combination of Bundle Adjustment with an evolutionary algorithm by means of 3D reconstruction of objects from visual information only.