Martin TONKO and Keisuke KINOSHITA
Error Description of Projectively Reconstructed Point Sets
Abstract:In this report, error propagation is derived for an existing sequence of equations, which allows to projectively reconstruct a 3D point set solely using given images of this point set. As input, image feature points, their position uncertainty, and the correspondence information must be given. All image feature points are assumed to be stochastically independent. The covariance tensor calculus is used to propagate the uncertainty of the input into an uncertainty of the output, which consists of projection matrices and 3D reconstructed points. In the course of error propagation all stochastic dependencies are modelled, resulting in an accurate description of the reconstruction error. This fact is experimentally validated. Furthermore, experiments with real image data are given.