3D reconstruction in a constrained camera system

Recently, there have been active researches on structure from discrete motion or shape from image sequences. In order to compute metric shapes, we may use a stereo algorithm with calibration parameters, or algorithms with a self-calibration method from three or more images. In this paper, we show a method of computing a shape from two images without using a calibration target. First, the camera system is constrained such that the intrinsic and the extrinsic parameters are limited to a minimum number. Then the constraints are relaxed step-by-step until real situations can be handled. The values of the parameters of the restricted system are used as the initial values to the more general configuration. This method can be used when we restrict the motion between two cameras. First, we derive equations with strict constraints, and later we relax the constraints to handle realistic situations. Experimental results with real images showed the validity of our algorithm. (C) 2002 Elsevier Science B.V. All rights reserved.