Abstract

Finding correspondences in the context of "wide-baseline stereo" is probably the last difficult missing step on the way to fully automatic reconstruction of 3D scenes from photographs. We propose to address this step from an object recognition point of view. Two new concepts are introduced. The concept of a singular region allows to detect a rich and stable set of objects of interest in each image. The concept of a measurement region allows us to derive for these objects discriminant features which are invariant to attqne image transforms, view-point changes, photometric changes, occlusion, etc.. We demonstrate our approach in a real experiment. A correct epipolar geometry between image pairs taken by an uncalibrated camera undergoing a large rotation, change of scale, viewpoint change, or occlusion is established fully automatically.

Citations