A robust global linear method for structure from motion

Abstract

A technique for building consistent 3D reconstructions from unordered large image sets based on a global linear method is presented. When views are treated incrementally, this external calibration can be subjected to drift, contrary to global methods that distribute residual errors evenly. We propose a combined global linear method based on computing consistent measurements in three views. First, all global camera rotations are computed from relative rotation estimates of pairwise image matches. Second, we minimize an approximate geometric error and projection error of feature points to find a linear relationship in camera triplets. This step can efficiently remove incorrect triplets which is very important for global reconstruction. Third, these triplets can be directly scaled up to register multiple cameras which can serve as a good initialization for final bundle adjustment. The performance of the proposed method is tested on several well-known image sets and the result is accurate and robust.