Maximum likelihood autocalibration

This paper addresses the problem of autocalibration, which is a critical step in existing uncalibrated structure from motion algorithms that utilize an initialization to avoid the local minima in metric bundle adjustment. Currently, all known direct (not non-linear) solutions to the uncalibrated structure from motion problem solve for a projective reconstruction that is related to metric by some unknown homography, and hence a necessary step in obtaining a metric reconstruction is the subsequent estimation of the rectifying homography, known as autocalibration. Although autocalibration is a well-studied problem, previous approaches have relied upon heuristic objective functions, and have a reputation for instability. We propose a maximum likelihood objective and show that it can be implemented robustly and efficiently and often provides substantially greater accuracy, especially when there are fewer views or greater noise.

Figure optionsDownload high-quality image (136 K)Download as PowerPoint slideHighlights
► Unbiased maximum likelihood objective for autocalibration.
► Monte carlo propagation of covariance.
► Instability of maximum a priori objective functions.
► Configuration dependency of optimal coefficient weighting.