Hierarchical feature grouping for stereo matching

We present a hierarchical approach for feature description and stereo matching in computer vision. Hierarchy is first used to make a high-level description of the scene. Edge pixels and lines are extracted and grouped into higher level symbolic descriptions: L-junctions, facets, surfaces. Because they are fewer in number and more significant, high-level features are easier to match, and help reduce matching ambiguity. Thus, the matching process is hierarchical: it starts with surfaces at the highest level, and is propagated down to lines at the lowest level. To determine the best order of matching, a control strategy is defined by classifying high-level features into areas of attention according to certainty and confidence criteria. Thus better formed features are matched first