Modified auto-calibration for 3D reconstruction from multiple views of an object

This paper explores the problem of Auto-calibration in 3-D reconstruction of an object from multiple images. Uncalibrated reconstruction, from a set of images taken from any ordinary CCD camera is complicated and difficult due to the presence of non-linear constraints on the scene. Any linear solution suffers from numerical instability or requires additional scene and camera information to be provided. We here present a method of 3-D reconstruction of an object which works without any additional camera parameters by modification of Auto-calibration procedure. The DIAC, computed by the Auto-calibration method is not always positive-definite and metric reconstruction cannot be performed. We approximate DIAC by its covariance square-root, guaranteeing its positive definiteness. This is compared with a few other closest positive definite solutions using a rigidity constraint, and is found to be quite satisfactory.