Stereo Without Search

European Conference on Computer Vision (ECCV) 1996 In its traditional formulation, stereo correspondence involves both searching and selecting. Given a feature in one scanline, the corresponding scanline in the other image is searched for the positions of similar features. Often more than one candidate is found, and the correct one must be selected. The problem of selection is unavoidable because di erent features look similar to each other. Search, on the other hand, is not inherent in the correspondence problem. We propose a representation of scanlines, called intrinsic curves, that avoids search over di erent disparities. The idea is to represent scanlines by means of local descriptor vectors, without regard for where in the image a descriptor is computed, but without losing information about the contiguity of image points. In fact, intrinsic curves are the paths that the descriptor vector traverses as an image scanline is traversed from left to right. Because the path in the space of descriptors ignores image position, intrinsic curves are invariant with respect to disparity under ideal circumstances. Establishing stereo correspondences is then reduced to the selection of one among few match candidates, a task simpli ed by the contiguity information carried by intrinsic curves. We analyze intrinsic curves both theoretically and for real images in the presence of noise, brightness bias, contrast uctuations, and moderate geometric distortion. We report preliminary experiments.