Combining Multi-view Stereo and Bundle Adjustment for Accurate Camera Calibration

Abstract

This article presents an algorithm to achieve accurate camera calibration for 3D reconstruction/visualization systems observing static scenes. The advent of high-resolution digital cameras, and sophisticated 3D reconstruction algorithms such as multi-view stereo offer the promise of unprecedented geometric fidelity in image-based modeling tasks, but it also puts unprecedented demands on camera calibration to fulfill these promises. Camera calibration is an essential step of most such systems involving multiple cameras. While there exist several standard procedure for the task, it is not easy to ensure accurate calibration. In this article, we talk about existing popular camera calibration procedure together with their problems and potential sources of errors, then provide a solution to these problems with an algorithm that produces accurate camera calibration starting from an initial guess possibly containing some errors. More concretely, the algorithm uses a multi-view stereo system on scaled-down input images to reconstruct rough 3D geometry of a scene from initial camera parameters, which is used to effectively guide the search for additional image correspondences. A standard bundle-adjustment algorithm is used with the obtained image correspondences to tighten-up camera calibration. The proposed method has been tested on various real datasets to prove its effectiveness.