By defining the disparity of a spherical stereo we reformulate the real-time stereo problem. Based upon this definition the realtime spherical stereo becomes a general model which can cope with cameras with any wide field of view including conventional ones. By transforming the rectified spherical images to latitudelongitude representation, the correspondence of feature points can be speeded up by the same processing as planar stereo images. The effectiveness of this approach is shown by realizing a real-time spherical stereo using a pair of fisheye cameras.