Can a one-eyed robot emulate mammalian binocular stereovision? Developments registered during the last two decades seem to support an affirmative answer. The present paper discusses the existing techniques for three-dimensional vision, focusing in particular on the rather difficult case of objects having curved surfaces.
Surface shape measurement can make use of two different techniques: shading and stereo imaging. The models that are used in image synthesis to shade an object can also be used to analyze an image consisting of gray levels in order to obtain the surface shape from the shading information, but the selection of the appropriate model (diffuse, specular, or mixed reflectors) is a very difficult problem.
The stereo technique is more accessible to mathematical manipulation (matrix transformation of coordinates) and is easily applicable to measurement by triangulation if the object presents polygonal faces and edges. For curved surfaces, the knack is to project a rectangular grid pattern; this enables the system, after calibration based on the three-dimensional coordinates of six points, to perform surface measurement using a single view of the object.
A special section is dedicated to the representation of curved surfaces, but the analysis is limited to quadrics and their reduction to ellipsoids, hyperboloids (with one or two sheets), paraboloids, and degenerate forms (cones, cylinders, planes, and straight lines). This section is simply a reproduction of classical results in analytical geometry, the unique novelty being the incorrect formulas (61) and (97), together with many other revision errors. The authors conclude this chapter by stating that “in cases where a surface is best described by other functions such as trigonometric, logarithmic, hyperbolic, etc., methods do exist [today] for which the exact model can be obtained.”
As a survey, the paper is superficial though correct in its description of the present state of the art, and incomplete in its presentation of the past evolution of this area, despite its 38 references. Nothing is said, in particular, on image processing in the brain, parallel processing techniques and collective computation in neuron-like circuits, or color vision. As a research contribution, it is poor in results and awkward in methodology. If the authors had intended to praise the intelligence and stimulate the interest of their potential readers, a complete and competent revision could have reduced the length of this paper from 50 pages to 10 or 15 in a more synthetic and careful presentation; this, regrettably, seems not to have been the case.