A multiple-baseline stereo for precise human face acquisition

Abstract

Stereo algorithms can provide useful three-dimensional information for a variety of purposes, including face recognition. For good accuracy, the distance between cameras (i.e. the baseline) should be wide. However, the large baseline requirement often makes the correspondence problem more difficult because (1) the region that must be searched to find correspondences is larger, increasing the frequency of ambiguous matches, (2) differences in viewing position lead to occlusions and window ditortions, i.e. regions that are not visible to all cameras and corresponding windows that do not represent the same surface patch, and (3) objects have non-Lambertian reflection characteristics.

Previous methods have addressed the problems of (1) and/or (2) by using multiple-baseline stereo pairs. However, because brightness difference minimization or other similar criteria is still used when searching for correspondences, problem (3) remains. In this paper, a refined multiple-baseline stereo technique is proposed, which estimates the brightness on each camera by using adjacent cameras thereby reducing changes in brightness due to non-Lambertian reflectance.

Experimental results are presented for images of a mannequin's face and significant reductions in correspondences errors are demonstrated in comparison with previous methods.