Calibrating a varying camera from silhouettes and background

In this thesis, we consider a specific 3D modeling scenario where a smooth 3D object of unknown shape is observed, against a background plane with unknown patterns, by an uncalibrated, moving camera with varying intrinsic parameters. Starting from the acquired images, we address the problem of recovering i) the value of the camera intrinsic parameters for each image, ii) the camera motion and, iii) the 3D reconstruction of some points on the object surface. The problem is particularly interesting in prosthetics, or custom cloth, shoe manufacturing, where the 3D model of a smooth textureless object is to be ob- tained. The smoothness of the object makes it difficult to identify image cor- respondences, due to the continuously varying contour generator as the camera viewpoint moves. In addition, using, e.g., consumer cameras (e.g., those provided with tablets or smart phones), often the auto-focus property of the device is ac- tive letting the intrinsic camera parameters vary between acquisitions. Therefore, calibration methods, such as Zhang [2000], cannot be applied. We propose a new framework that could calibrate a fully varying camera from background and silhouettes of smooth objects, not requiring more than two frontier points per view-pair.The epipole positions in each image pair are first estimated by finding two epipolar lines that are tangent to the object silhouette with the help of the plane induced homography. The projective reconstruction of the im- age sequence is then robustly computed from the estimated epipolar geometry. A flexible self-calibration algorithm using the absolute dual quadric is employed to determine the camera intrinsic parameters and to upgrade the projective recon- struction to metric. The proposed algorithm does not depend on the assumption of orthographic or affine cameras to find epipoles thus is applicable to generic projective cameras with both varying focal length and principal point. Experiments with both synthetic and real images are carried out and good calibration accuracy is achieved.

see English version