A global camera network calibration method with Linear Programming

We propose a novel global calibration method for a network of cameras. Given a set of unknown cameras linked by epipolar geometry, we transform it into a graph of camera triplets. Underlying this new graph is a set of trifocal tensors. Based on (i) a linear computation of the trifocal tensor given two of the three related fundamental matrices, and (ii) considering a maximal tree embedded in our graph, we first design a way to globally recover the cameras. Then, we observe that considering the whole graph consists in simply adding constraints that can be easily linearized. A global estimation of the cameras thus boils down to solving a linear program. Numerical experiments show the ability of our method to recover accurate geometries, without any incremental process, and dealing naturally with loops in the network of cameras. Moreover, the number of unknown parameters is limited: we only consider the cameras (the motion), not the three-dimensional points (the structure).