Enhanced surface reconstruction from wide baseline images

Abstract

This work deals with the problem of dense matching from stereo images under wide baseline conditions. By considering the characteristic properties of widely separated views, we propose an extension to a recently published algorithm for detailed reconstruction of continuous surfaces. The algorithm compensates for the occurrent affine distortion between the views to allow low level intensity based comparison necessary for dense matching. The matching itself is performed by an enhanced region rowing based affine propagation method that takes surface distortion into account to handle complex piecewise-smooth surfaces. It is experimentally shown that this new method can achieve smooth and accurate reconstruction from wide baseline images of both indoor and outdoor scenes. To quantify the reconstruction results we have created realistic synthetic datasets with ground truth. These datasets form the core of a future testbed for comparison of different wide baseline surface reconstruction techniques. In the current study, results on the synthetic images are compared to the ground truth to measure the accuracy of our method.