This paper presents an effective method of selecting candidates for correspondence in segment-based stereo. Boundaries of objects are extracted from stereo images and segmented at feature points. The resulting line or curve segments are used as units for correspondence. The initial candidates for correspondence are obtained based on the epipolar constraint. And, more reliable candidates are selected using the features of each segment such as intensity, direction, and shape. When there is another camera in addition to the basic two cameras, the image can be used to further reduce the candidates.
