Figure 9 Experimental Results on Real Images 3.3 Region Inference 4 Experimental Results 3.1 Saliency Tensor

on the Renault part scene. Illustrated in the shaded view of the scene description is the inferred region for the half-oc-cluded background (compare to the correspondence data). A rectified, texture mapped view of the scene is also presented. Notice that the left side of the Renault part, which is mostly occluded, is correctly inferred. In both texture mapped views, inferred regions with no texture information are given random texture. We also applied our algorithm to a building scene captured by aerial image pair, depicted in figure 10. Using the knowledge that the target object is block-like building, we combine edge information with the inferred overlapping roof surfaces to derive vertical surfaces that are visible in both images. Inference of vertical surfaces is hard as they are often half occluded, or difficult to obtain by local correlation measurements. Also note that surfaces that are too small to provide correct correspondence are not detected. 5 Conclusion We have presented a surface from stereo method which addresses both the correspondence problem and the surface reconstruction problem simultaneously by directly extracting scene description from local measurements of point and line correspondences. Unlike most previous approaches , it explicitly addresses the occlusion process, leading to overlapping surface descriptions, that is multiple depth values for a given pixel. The method is not iterative, the only free parameter is scale, which was kept constant for all experiments shown here. We hope to demonstrate that this approach is also applicable to transparent surfaces with no change. A Bayesian treatment of the stereo correspondence problem using half occluded regions " , (a) a rectified, texture mapped view of the book scene (b) the Renault part scene Figure 10 Result on a building scene 6 junction curves and junction points can be extracted by a non-maximal suppression process [13] modified from the marching process [8]. Figure 6 depicts a slice of the inferred surface saliency for the book example. Note that although we use a specific surface model in our estimation, the estimation errors due to model misfit are incorporated as orientation uncertainties at all locations and are absorbed in the non-maximal suppression process. While we only presence our framework in the context of inferring surfaces, this tensor voting technique can also be applied to inferring curves, simply by changing the interpretation of the tensor components and the voting sa-liency tensor fields. In this section, we address the problem of …