Quasi-perspective stereo-motion for 3D reconstruction

Abstract

One important motivation of integrating stereo vision and visual motion into the so-called stereo-motion cue is to make the two original vision cues complementary, in the sense that (i) the ease of establishing motion correspondences and (ii) the accuracy of 3D reconstruction under stereo vision can be put together for bypassing or overcoming (i) the generally difficult stereo correspondence problem and (ii) the limited reconstruction accuracy of the motion cue. The objective is to allow a relatively short stereo pair of videos to be adequate for recovering accurate 3D information. A previous work has addressed the issue, which lets the easily acquirable motion correspondences be used to infer the stereo correspondences. Yet the inference mechanism requires to assume the affine projection model of the cameras. This work further extends from the affine camera assumption to quasi-perspective projection models of cameras. A novel stereo-motion model under quasi-perspective projection is proposed, and a simple and fast 3D reconstruction algorithm is given. Only a small number of stereo correspondences are required for reconstruction. Experimental results on real image data are shown to demonstrate the effectiveness of the mechanism.