Reconstructing visible surfaces

Abstract
For many high-level vision tasks, such as scene interpretation or object recognition, it is advantageous to know the visible surface of the object space. An approach based on scale space techniques and matching in object space is described. In every discrete step in the scale space images are matched which are warped to the surface obtained in the previous step. This corresponds conceptually to matching in object space, and it helps reduce the foreshortening problems that are associated with any matching method. The 3-D positions of the matched points form a sparse set that needs to be densified to obtain the surface. On every level the matching results are analyzed and a hypothesis about breaklines and occlusions is formulated. The surface is found by interpolating the 3-D points found by matching. The interpolation stops at boundaries of suspected breaklines and occlusions. An independent analysis of the surface normals may confirm or reject the hypothesis. Yet another independent clue about breaklines or occlusions stems from edges. At the final step the warped images become orthophotos and the matching vector vanishes for all points. The theoretical description of the model is followed by experimental results.

This publication has 0 references indexed in Scilit: