From: AAAI-82 Proceedings. Copyright ©1982, AAAI (www.aaai.org). All rights reserved. LOCAL COMPUTATION OF SHAPE Alex Paul Pentlsnd Artificial SRI International, Intelligence 333 Ravenswood Center Ave, Menlo Park, CA. 94025 ABSTRACT Local, to infer parallel scene. of shading shape without surface viewed analysis The able assumption following about second derivative ponent of surface orientation maximum-likelihood estimate of orientation compute surface without knowledge been to be determined natural comand a exactly, for each image shape fashion from local image then it should information take possibility. we examine an image, usually in shading (changes a small neighborhood in image intensity). of a particular is small changes Finding a contour point is unusual. image if we are to investigate what we may learn about point from local examination of an image, images, has with shading. and surface Thus shape Thus shape ourselves the question: from an unfamiliar within surface we must concern this paper examines be recovered local analysis a point around all we find in the neighborhood the neighborhood of shape. in full advantage to This algorithm and synthesized calculate bottom-up When of the has been developed in parallel, that a reason- depth component of scene characteristics. on both Given direct, the allows the image-plane An algorithm a good estimate about local analysis of the remaining orientation, evaluated produces smoothness, of image intensity to be made. knowledge has been proven: surface can possibly in an image may be used o priori how may image by using of shading? II. Previous Work I. Introduction Horn and his colleagues A spatially the restricted first stage relatively because of any visual simple, quite complex. for the remainder of analysis produces remainder of the visual system directly with Determining as much about is logically analysis might that of the system. will be much then simpler all of the ambiguities than the if it If a visual known in which the illumination is known Support for the Army Research 05ce contract DAAG2479-C-0210. portion of this research was substantial Artificial Intelligence of Technology. in part by Department stages Laboratory Support the of Defence of this research was accomplished of the Massachusetts for the laboratory’s Advanced N00014-75-C-0643 Grant final Research under 05ce research Project at the techniques information; face whose shape of the an initial shape-from-shading there the image, tech- is su5cient such as in a fac- and the surface reflectance theory shape-from-shading of the scene, it fails to answer the quesshape in an unfamiliar satisfy image. our requirement constraint by smooth from occluding is to be estimated. Thus only techniques boundary contours) Further, of using Horn’s shape-from-shading by propagating as provided along conditions over the sur- further analysis required. is provided Agency Science (such Institute of Naval Research and in part by National function A about Horn’s knowledge tion of bow to determine provided function orientation beforehand. because sumes a priori local image by a priori of (1) the distribution for which information none of these l in situations to solve reflectance These surface. may be useful However, system knowledge and (3) the surface on the object’s several for using image intensity (2) the bidirectional surface, function is therefore schemes tory setting of the image. the world as is possible at this first stage of processing. shape given D priori niques the If the first stage of the world, for object curve to the determines integration object’s [l], [2]) have developed (e.g., numerical of illumination, important is available and, therefore, be or it might be is especially a rich description had to deal important This the information system, image image intensity, This first stage of analysis of the visual requirements of a single system. e.g., measuring it determines remainder analysis of the III. The Estimation contract Of Surface Orientation Foundation The MCS77-07569. mation 22 problem of estimating is, essentially, surface tbe problem shape from local of determining infor- the unknown surface normal, N, from image measurements I, or its derivatives. equations Solving requires Because having the surface require of image intensity, for the unknowns more normal in any system measurements than N has two degrees of unknowns. of freedom, we at each image point to obtain least two measurements a solution. Therefore, we cannot image intensity alone, ment per image point. obtained from intensity; ber of unknowns for N simply per point; or higher each per point, to solve per point - does Therefore, to obtain solve more measurements for N locally, we must make some of unknowns -_ . _.._ Figure 1 The manner in whrch Image curvature %preads” indicates the tilt of the surface. This may he understood by imagining that we could observe the lines of curvature on a surface directly. These lines of curvature would look just like the lines drawn in this figure. If we were looking straight down on a surface with no twist, the lines of curvature would appear perpendicular, as in (a). As we tilted the surface off to one side, the lines of curvature would appear progressively more spread, as in (b) and (c). Different directions of tilt cause spreading in different directions. give we cannot C B A can be of image derivative so that the number larger that the number using measure- derivatives additional into the equations. assumptions normal one it also brings an even larger num- by using derivatives in order simplifying second, while more measurements the surface it provides only More measurements the first, however, determine because is not available to us. of measurements Thus one of the two components A. The Tilt tilt, may be determined Of The Surface tensity When we observe impression the a smooth surface away from slants impression be possible we could determine component the tilt of freedom orientation) would the of the (slant, the lines of curvature would we were looking straight lines of curvature (a). surface appear (c). Different down appear the surface progressively directions This leaves is a second-order orientation, tilt directly. exactly, then surface only surface. in N. on a surface directly. on a surface with perpendicular, “twisting” 1. If no twist, the error the surface 1 as in Figure of tilt would cause spreading typically causes between where errors on the order points across on the a surface, the is on the order of The largest errors occur for ljpUVjj > > 0. For such surfaces, those with of l/lOOth the assumption surface f(u, v) of an arbitrary observed by this approximation can he on the order second-order to one side, the lines of curvature more spread, surfaces, [3]. Thus we are observing + ve:! + f(u, v)ef, l/lOOOth of tbe width of the surface. These as in Figure uel if ten p0int.s are observed error incurred in- direction. to he determined. that the surface 6 is the spacing Thus maximum that we of the illuminant + ozv ‘. The approximation a surface by such of 63, where the of image that is, a surface that may be described p = patch is of the form*01u2 of surface the tilt? Imagine assumes surface, by the Monge If knowledge orientation, derivative only the slant of the surface This proposition a strong look like the lines drawn in Figure would As we tilted would of surface surface be left undetermined lines of curvature we have the depth-component How we might go about estimating could observe a strong direction Because to compute one degree - us. of the image-plane it might we obtain that is, which surface, of the tilt of the surface without directly of surface from the second of the width that we are observing does not introduce of the such a much error. 1 (b) and B. The in different Slant Surface Of The directions. We can not observe of course, lines of curvature but we can observe with the illuminant The second the interaction in the second derivative of image Pentland on the surface directly, of surface derivatives intensity curvature completely of image intensity. may has three components: spread tilt. of the lines of curvature, term is the greatest The direction direction along the direction is also the direction in which the spread which @I in which this of a smooth, Lz#Iy,, direction sity, 81, I zy #0, Given recond-order rurface then the tilt of the surface in which the second derivative estimate surface, with thogonal n, direction, of tbe surface slant, be We as Lsplacian of a smooth, Of Image homogeneous K2,z7;2 - tc?, K,, is the surface and cannot of the surface. then VI -=I direction the tilt of the surface of the surface propositions. an image is also the an image while (Normalized Given and so the following of the Surface) homogeneous an unbiased second-order where (Tilt that the slant from the curvature in the following Proposition proposition: Proposition make Intensity) of the surface is the greatest is the greatest, still disentangled developed Lz and I,,, the %urvaturen of image intensity along the z and y axes, and Izv, the ‘spread* of those curvatures. Just as with the spread [4] has proven may be exacctly determined, curvature along the rurface is the surface curvature t‘~ is the t component tilt in the or- of the surface is the image of image inten- *Linear is greatest. and appropriately 23 constant positioning terms in f may the coordinate be accounted axes. for by normal (which is equal to the arc cooie of the surface slant) What this proposition shows only of the squared curvatures the foreshortening. Note the surface albedo The range and the V21/1 that of the surface curvature say a uniform distribution, had a particular uniform face likelihood distribution, value of V*1/1, estimate Of of the foreshort of ZN (the equal to the Slant) to zE2, we Assuming curvature a the msximum- t component arccosine of the of the slant our- of the is /I V21 where of proposition. of surface estimate surface) a priori is proportional (Estimation normal, value of the slant of the surface, i.e., COS-~(ZN). distribution likelihood curvatures observed then for any observed This leads to the following Proposition the surface particular the foreshortening then have an eArnate and if we were given that the magnitude we could make a maximum Because between for any Therefore, nz and n?,, and of the illuminant in this quantity. of relationships is limited. ening. the effects do not appear foreshortening V21/1 is a function is that of the surface, a: ia the variance I xi of the distribution of surface curvatures. Pentland degree [4] and Bruss of freedom information. The parameters leaving likelihood estimate ehading the tilt estimate together one of the two neither knowledge Figure 2 Evaluation on synthetic images. (A) Artificial images of a sphere and an ellipsoidal shape used to test the surface shape estimation algorithm, (B) side view of relief maps showing the true shape. (C) Relief maps showing the estimate of surface shape made by the shape algorithm for the sphere and ellipsoidal shape images. Compare these relief maps to those of 4 (B). exactly, gives by definition Therefore, constitute one shading proposition possible Note that require This estimate. it is theoretically information. that of the slant, which unbiased at least by the can be determined undetermined. and the tilt propositions that showed orientation the slant that undetermined tilt proposition the minimum-variance of surface [S] demonstrate remain of surface only maximum will a is &ant the the 6errt e&mate to make from local the slant estimate of the illuminant A. nor Synthetic direction. The shape generated On the basis of this theory, M.I.T.‘s synt,hetic LISP machines.*This and natural images. the level of performance under ideal conditions, the performance the more varied, scenes. of the shape complex Some examples on both synthetic Using algorithm images algorithm of the application which on both allows images to be evaluated was appear tested timated surface quatcly evaluat,e the shape allows tegrated under maps *The algorithm is a straightforward implementation in the slant and tilt propositions the 2 (A). do not the performance algorithms’ to produce were found the estimates a relief shape, presented The of the cal- algorithm’s above. 24 relief map estimate which that allow of the an observer results es- to ade- Therefore, orientation an adequate were in- These relief impression are the output in this paper even though of surface surface of the algorithm, map of the surface. and so they the of the of the algorithm. of surface 2 (B) showing displays performance to give observers surface for the examples here. orientation syntheticly Figure this image. estimates however, of displaying the estimated of this shape algorithm produces found, is not part of the shape algorithm culations using in Figure was used to generate algorithm it was for purposes found in natural scenes are presented which shape orientation; to be checked the use of natural shape The the on one of has been tested synthet,ic and noisy conditions and natural for estimating was implemented of the shape whereas surface an algorithm algorithm algorithm images shows the side view of a relief map for this surface, IV. Evaluation surface shape (the “shape algorithm”) Images of shown integration per ue. from orientation integrating is shown the shape in Figure 2 (C). Comparing the estimate the relief map which of surface orientation 2 (B) s h ows that the shape map in Figure a very high level of accuracy. sphere, 81 using convolutions It is important knowledge shape was recovered map generated orientation. image orientation. (A) and 3 (B) correspond shapes of these his impression The relief maps in Figure as can be determined two objects. algorithm on other natural The reader should 3 to the compare images with the relief has also been images, to the actual parallel, of surface surface local does not require successfully and on electron case the estimates sufficient of a rock, 3 (A) and 3 (B). shape the of surface image from the shape algorithms The as closely Figure 3 with the relief estimates of shape from the digitized maps of Figure which algorithms 3 (B) s h ows the digitized of surface here. of a log, together wit.h the relief map generated closely in any a priori on several natural will be presented from the shape Figure estimates that without has also been tested (A) shows the digitized each 0.01 %. Images and two such examples actual to within that these shape were calculated local image measures, The shape algorithm together can attain of scene characteristics. B. Natural images, “truen relief algorithm X 200 pixel image of a For a 200 to remember from purely from integrating with 21 X 21 pixel masks to calculate and V21, the correct parallel results to the original shape shape. images. produced These examples computation employed microscope described any a priori knowledge to obt*ain a useful estimate of surface In correspond demonstrate in this paper, of the scene, is 1Figure 3 Evaluation on natural images. (A) The digitized log image __ _._ _ . . and the relief map generated by the shape algorithm for that. image. (El) The digitized rock image and the relief map generated by the shape algorithm for that image. shape in natural images. REFERENCES [l] Horn, B.K.P.,“Shape From Shading: the Shape of a Smooth Technical Report of Technology, [2] Ikeuchi, Shading K. 79, Project Special Horn, E., *and Keller, A.P.,“The Massachusetts B.K.P.,“NumericaI Vision, A.I. Institute Shape Artificial from Intelligence, 15 (1981) H. B., “Analysis Visual Inference From Local Features,” Massachusetts A., Thesis, Institute for Obtaining from One View”, Of Numerical (John Wiley and Sons, New York, New York, 1966) [4] Pentland, [S] Bruss, MAC, Boundaries,” Issue on Computer Methods,” A Method Object (1970) and and Occluding [3] Issacson, Ph.D. Opaque Ph.D. Thesis, Institute “Shape Dept. of Technology From Shading Elec. of Technology, Engr.& Of Shape: Computation Psychology Department, (1982) And Bounding Comp. Contour,” Sci., Massachusetts (1981). 25