Linear Image Features in Stereopsis

Proceedings of the AAAI Conference on Artificial Intelligence, 5

Volume

Science

Perception and Robotics

Most proposed algorithms for solving the stereo correspondence problem have used matching based in some way on linear image features. Here the geometric effect of a change in viewing position on the output of a linear filter is modeled. A simple local computation is shown to provide confidence intervals for the difference between filter outputs at corresponding points. Examples of the use of the confidence interval are provided. For some widely used filters, the confidence intervals are tightest at isolated vertical step edges, lending support to the idea of using edge-like features in stereopsis. However, the same conclusion does not apply to image regions with more complicated variation on the scale of the filter support.