Improvement of Multi-View Stereo Vision by SEVL

Abstract

The latest stereo algorithms use multiple-baseline stereo with many cameras to make sure of stability and to cope with occlusion. But the increase of stereo pairs still has difficulty in estimating true correspondence. In this paper we propose an improvement method for multiple-baseline stereo by SEVL (Stereo with Evaluation Value Limiter). The SEVL is a correspondence method that limits evaluation values. This method can avoid the bad effects of incorrectly corresponded stereo pairs and can cope with the occlusion without the stereo-pair selection as conventional occlusion avoidable methods. We first explain that false correspondence in occluded area by multiple-baseline stereo is caused by bad effects of evaluation values generated by incorrectly corresponded stereo pairs. We next show the bad effects of such stereo pairs can be restrained by SEVL. We compare the SEVL to other occlusion avoidable methods and show the superiority of SEVL. We have confirmed that the combination of these two methods has superiority to each single method.