Anda di halaman 1dari 8
An Efficiently Computable Metric for Comparing Polygonal Shapes Esther M. Arkin, L. Paul Chew, Daniel P. Huttenlocher, Klara Kedem, and Joseph S. B. Mitchell Aburaci— Mode-bsed recogiin i concerned with comparing 4 shape A whieh i stored ax mode! for wome particular bjt, wih 1thape 8, which is band to ext i an image A ad Ba cle So Being the sae shape, then vison ‘ould be reasonably ey to compa and’ should atch oa intone asers shoul besa fo thone hat person might give We develop a meted for comparing pogo that has hese properties The method is ased on the Ly ditance betwen the turn Tuntins ‘tte to palygon: works for both cones and tonctenesPlptons ‘tnd'ras i ime Om log me) here the umber of eres io Index Terms—Computationsl geometry, stance meee, model-based matching. shape cimparn,siarty transformation, trang og (theta representation, 1. Iwrroouenon A SROBLEM of bah hoe and practical imponance in computer vision is that of comparing two shapes. To what extent is shape A similar to shape #7 Model-based recognition is concerned with comparing a shape A, which is sored as a ‘model for some particular object, with shape B, which is found to exist in an image. IA and Bare close to being of the same shape, then a vision system should report a match and return a measure of how good that match is. Hence, we are iteested in efining and computing a cos function dd, B) associated with two shapes 4 and B that measures their dissimilarity ‘The long-ierm goal ofthis research isto develop methods of ‘comparing atbitary shapes in two of thee dimensions, Here, \we restrict our attention to polygons! shapes in the plane, with an extension 10 the cise in which a boundary may contain circular ares in addition to straight line segments. Our technique is designed to work with objec fr which the entre boundaries ate known, ‘Before suggesting @ measure to be used for comparing poly- ‘gons, we examine several properties that such a measure a, Massp ciel 17, 1989 vied Oc 1980, Recommended {or acepane by O-T. Teast This work was suppres tn pr bythe ‘ies! Science Founaton under Grants DMC 84S See BOSE BST 2nd DMs 8500508 LF. Chew and K- Redo wee sapped by DARPA Shcer ONR Const NOMAGK-SHL NSF Grant BMC Ae 17386 a OONK Contact NOOOA6-K-AD81 15-8" Atal wan upped In pt by the Nana Scene Footie veder Grants IRLATHINGS nd BOSE 'Skst6< a bya gra fom Hughes Resear Labor TEM Arkin and15 8 Michel re wath Sebo of Operstions Resch nd ints Entering, Carell Une, Hey, NY 18S LP Chew and D.P. Hutetlocter ar wih the Departmen of Comput Sckee,Comell Une Iaea NY 1853. Kode was lh te Departmen of Campeter Scene, Corel Univer sy es NY T4883. She snow wth te Deparment of Comput Sine, {al See Une, Tel A aa TEEE Log Number 905%, should have (for related arguments see {7). + Tt should be a met, (A,B) > 0 forall A and B. (A, B) = 0 i and only iA = B, We expecta shape 0 resemble ite (A,B) = d{B, A) for all A and B (Symmetry). The onder ‘of comparison should not matter (A,B) + d(B,C) > d{A,C) forall A, B, and C (Teiangle Inequality), ‘The triangle inequality is necessary since without it we can have @ ease in which d(A,B) and d(B,C) are both very small, but d(A,C) is very large. This is undesirable for patlrn matching and visual recognition applications. 1 is very similar to B and B is very similar tC, then A and C Should not be too disse, + Itshould be invariant under translation, rotation, and change- ‘ofseal. In other words, we want 19 measore shape alone + Te shouldbe reasonably cay to comput. This must hold for the measute to be of practical use + Most important of all, it should match our intuitive notions fof shape resemblance. In other words, answers should be Similar to thore that a human might give. In particular, the measure shouldbe insensitive to small perturbations (ot smal eos) in the data, For example, moving a vertex by 4 small amount or breaking a single edge into two edges Should not have a large effect. A. Representation of Polygons A standard method of representing « simple polygon A is to describe its boundary by giving a (circular list of vertices, expressing each vertex as a coordinate pair, An allemative representation of the boundary ofa simple polygon 4 isto give the turning function © (3). The fapetion O48) measures the angle ofthe counterclockwise tangent as a fonction of the ae length , measuted from some reference point O on A's boundary. Thus (0) is the angle v that the tangent at the reference point O makes with some reference orientation associated with the polygon (Such as the x-axis). @,(s) keeps track of the turning that takes place, increasing with lefchand turns and decreasing with right-hand tues (ee Fig. 1). Formally if m(3) is the curvature function fora curve, then &(s) = 6"(s) [13]. The curvature function n(s) is frequently used asa shape signature (5), (6), (11), (14. ‘Other authors ave used a slightly diferent definition of the turning function in which (0) is defined to be 0. Ove definition, in which ©4(0) is the angle of the tangent line at the reference pont, leads to a simple correspondence between 8 shift of © (5) in the # direction and a rotation of A. This correspondence is less elear fo the aerate definition. Without oss of generality, we assume that each polygon is rescaled so thatthe total perimeter length is Is hence, Oy is 2 162-s824914500-020050100 © 1991 IEEE 0 TEBE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE IRTELLIGENCE, VOL. 13, NO. 9, MARCHE 18H a pee ae Fa. Defiing the tr fancon (2) function from [0,1] 1 R. For 2 convex polygon A, O,(s) is 4 monotone function, starting al some value v and increasing tov + 2x. For a nonconvex polygon, (3) may become arbitrarily large, since it accumulates the tua! amount of tun, which can grow as a polygon “spirals” inward. Although © (8) may become very large over the intervals © [0,1] in order for the funeton fo represent a simple closed curve, we must have x(1) = ©4(0) + 2r (assuming thatthe origin O is placed at a differentiable point along the curve). ‘The domain of ©.) can be extended tothe entire real line Jn natural way by allowing angles to continue to accumulate as we continue around the perimeter ofthe polygon A. Thus, for 8 simple closed polygon, the value of @4(#-+1) is @x(s) 427 for alls. Note thatthe function ©.) i well-defined even for arbitrary (aot necessarily simple or closed or polygonal) paths A fn the plane. When the path is polygonal, the turing function is piecewise-constant, with jump points corresponding 10 the vertices of A. Representation of planar curves (an, in particular, polygons) in tems of some function of arc length has been used by a ‘numberof othe researchers in computational geometry ( [8 [11) and computer vision (ef. [2) We use this representation to ‘compute distance function for comparing two simple polygons (A and B) by looking at natural notions of distances between the Turning functions ©,(s) and © p(s}. "The function Q,(s) has several properties which make it especialy suitable for our purposes. I is piecewise-constant for polygons (and polygonal paths), making computations paricu- Jary easy and fast. By definition, the function © 4(3) is invariant under translation and sealing of the polygon A. Rotation of A corresponds to simple shift of (3) inthe @ direction. Note sso that changing the loation of the origin O by an amount € [01] along the perimeter of polygon A coresponds to horizontal shift of the function © (8) and is simple to compute [the now tuning function is piven by @a(# +8) AB Fig. 2. Nonaiform ut able for he roped tance ection We formally define the distance function between two poly- gont A-and B a8 the L, distance between their two turning functions © 4(e) and Of (s}, minimized with respect to vertical and horizontal shits of the turing functions (in other words, we 2 we show tat his distance function canbe computed cffciemly in O(n log) time for polygons with m vertices. ‘One possible drawback of our distance function i that it may ‘be unstbl er certain kind of olsen parla, nonuniform rose, For example suppose we havea angle with one Very ‘vagy side such a the one shown i Fig. 2. Comparing i tora triangle we wil get avery bad (fact abivarly bad) mach, because the proportion of the perimeter eoresponding tothe sides labeled A and B can approach 2c. Fortunately, 1 ‘many computer vision applications reasonable to assume that the noite oughly wniformly distbuted over the sides ofthe plygon, in which ase the similaiy measure we define performs cey, See Section IV for examples. ‘Schwarte and Shani (11] have defined 2 notion of distance similar to our. However, they compute an approximation based fon dicretring the trong functions ofthe to shapes ito many ually spaced pointy this, the quali ofthe approximation fepends on the number of poins chosen. Our approach, on the other hand, i (o examine the combinatorial complexity ‘of computing the exact metic funtion between two polygon boundaries, sing only the orginal vertices. Thus, our method rune in ime O(n logn) where nis the total numberof polygon ‘vertices, wile their method computes an approximate distance in ime O(log), where £>> ni the total numberof inte polation points uscd. Furthermore, [11] sbgget“convexitng”™ onconvex polygons in order to compare them, a8 their method doesnot apply to nonconvex polygons. Our algorithm apes beth convex and nonconvex polygons (and evento nowsimple polygons) ‘The remainder of this paper is organized as follows. In Section It we give formal dfiton of he dsance between two polygons ased on tei tring functions. We prove that this function i a metic and show some fis properties. These results are used in Section I to develop an OCW) algorithm for Computing the distance between two polygons, where i th {oul number of vertices: we then reine this algrithn obt fn O(n'logn) runing time. Section IV contins examples of the distance function computed for several polygons using 31 implementation of our method. Section V i a summary and 0 Troe ley (everywhere por, symm and has the Hotty opr, eons I le 4 wee HSS sete properies Ne sow show th ds) abo obeys the Single eggs 4s(80B) GRC) S Gchc) oy a Staightorwardaplicon ofthe Minkowal incully for Zy itr and ee mine (AB) be dB) = [Ds "(ata)". Sima let and be the nimi St dg(B0) We de = tact hand Bg Now GB gac) [fretesea-oun sera) +[fiesie+so 209 ena [fewest sates nse + [fteaterna ect 4 suatas|” “which bythe Minkowski inequaliy (ef (9) is 2 [fiesta ri) -Olerh tte +044) 000 0-ra -[[1e.0409-2000 +r = DESW.8), om ig. 3. The rectnglar stip famed bythe fonctions (6) ané ®(s). Clearly DS°(e,0") > dy( ALC) becanse he Later i minimized overall vlus of and incding and # 0 “Lemme 2: For any fixed value. off, and for any p> 1, DAME) in a conver function of 6 Proof, For fixed s and fixed t, the function F(@) = |@a(s 8) — @n(a) + 8 is cesly@ convex fuetion by the sonvexity of Gly) ~ jl? (or p= 1) and imtegrating 2 convex function maitains convexty (Ley if 4 postive vad function ‘ye i conven n'y for fined, then J y(zsy) de i also 8 onvex onetion of 9) a Tn particular, using the Ly metric, DS" (t,8) is @ quadratic funcxon of @ foray fixed vale oft. This Bas fo ay closed shapes, ut it i especialy easy 060 for poy gos. ‘We assume from ow on that A and Bae polygons. Then, for a fined fhe integral f",(«4£) ~ Gala) + Ode ean te computed by adding up the value ofthe integral within each Sirip defined by a consecutive pat of discontnutis in © (3) and Oye) (ee Fig 3). Tho integral within sip toa Computed a the width ofthe sip times the square of the Sitereace 6 (0-+8)~ O(a) (Bich fe onsant within each Sti), Nowe tht fm an ae the numbers of the veres In ‘Aan B, respectively, then there am warps and th 8 9 Changes, the value ofthe integral fr eich stip is quale function of @ In oner to compute dA, 2), we must minimize Di *(¢,8) cover all and 8. We begin fading the optimal # fr any Hed Sale of To simplify notation in the following dscuesion, we te fa) = Ba(o). ae) = Pal) and le 8) = Deo) Lemna 3: Let M8) =f (Jla-+1) — 98) + fds. The, in conde 0 minimize t,8), the bet value of @ is given by wins [worse ore ant, P= [a9 ajie+ ~dais)de a2 [lor 0— alate ‘Lemma 2 assures us that the minimum occurs when we Set this quantity equal to zero and solve for 8. Thus, vw J 009)- Ho 0a Now, fy Sls t)ds [oto =f tevdes [lo +2eies [soydes [soyderree act [Hedds Thus, wom [ aaron f peyae a2, a Substituting the expression for (4) in ds(A, B) we are left with a one-variable minimization problem, 1 (4,2) = {nie neocon} = {op | Ue+9-ser-woorsare e+ ana}! fa = {i [fuer ~oitas- cor]. ML Atconmane Derans In this section we show tht the distance function achieves its rinimum at one of mn discrete points on (0,1), which we rca events Reeal that nthe proces of Banding (4, 8) we have to shift the fonetion (9) f(s + 1) for (0,1, During this shiting operation the breakpoint of folie with the breakpoins fof g. We defines erica event a vale of rere a break of fcolides with a breakpoint of g. Clealy thre are mn such erica events form breakpoints i f and breakpoints ing Using the fat thatthe minimum x obaied at rts vent, ‘we present basic algoritin for computing dy, Bhat usin ‘O(r) time foro nvertex polygons (end ime Oran + )) for an m vertex polygon and en vertex polygon). We then sere howto moi the base method to fmprove the runtime to Ofn*togn) (or Om log mn for unegeal numbers of vertices) Recall that di(A, B) = ming» (h(t,8))4, where (t,8) = Dz-?(t,8). We prove that h(t) bas propentics that lead to efficient algorithms for computing ov polygon metic. “Lemma # It f() and 9) ae two piesewse constant fonctions ‘ith mand m breakpoints, respectively, then for contat 8 108) [wera 40h {is pecewise-linar asa funtion of, with mn breakpoints which sae independent of the value 8. Proof: We give & geometic proof. First recall that for a siven value of the discontinuities in f and g define a set of imeen rectangular sips (See Fig. 3) The value of A(t,8) is Simply the sum overall these strips ofthe width ofa strip times the square of is height. Except at ertial events, as fis shifted the width of each strip changes, Dut the eight remains constant Each changing rectangle contributes to changes in h(t, ¢)- If cis the amount of sift then fora shrinking rectangle, the change is (Cd times the square ofthe height; fora growing rectangle the change i (+2 times the square ofthe eight Since the heighs| fe constant, the change in (8) is a sum of linear terms and is therefore linear. Breakpoins in h(t,2) clearly occur at each of the mon ential events where discontinuity of fis aligned with a discontinuity of g a "This esl! leads toa staightforvard algorithm for computing 1&(A,B). Let (0°) be the location ofthe minimum value of 1it,6). By the preceding lems, h(t, 8") is piecewiseinear a5 fonction of with breakpoinss among a fixed set of ertical walues; thus, * must be at oge of the eiical values, Now, We,0°(0) = ACEO) — [OE = A(e.0) — [a — 27} (fom (A, 50 sutces to evaluate A(t) = f [f(e-+ 2) ~ 9a)ds 1 cial values of Corolary $: The distance ds(A, B) between two polygons A and B (with m and vertices) can be computed exactly in time Ofmnm + n)) Proof Forgiven values oft and 0, h(t, 8) can be computed in O(m-+n) time by adding the contributions of the m+n rectangular stripe between f and. Lel cis Cpn BE IDE critical events that occur as fis shifted by ¥. By the preceding observations, the minimum of At,@) occurs when t equals fone Of €y,¢j--") Emme Since the best @ valve for a given 1 [namely, #"()] can be found in constant time (Lemms 3}, we simply compute h(t,0"()) in Ofm + n) time foreach ofthese crits events, find the minimum, and tke its square rot 10 gel (A,B). a A. Refinement of the Algorithm ‘Te above time bound can be improved by using «somewhat more complex algoritim Theorem 6: The distance dy(A,B) between two polygons A and B (with m and m vertices) can be computed exaey fn ine Ofmlogmn). ‘Proof We prove the theorem by describing the algorithm. The basic idea's the same as the previous algorithm: we compute (9° (0) for each ofthe eral valves oft. By the commen before Coralay 5, it sufces 10 evaluate A, 0) LUS(3-+0)~ o(9)Feds at etal valves of Now we observe ihat (0) vases with «in a very constrained fashion. AS & rater of fact, by Reping rack of 4 sal tof values we an easily determine how the fonction f(,0) ehanges at each eritcal even. “The values we Keep ack of are based om the rectangular strips that appear between the two fnetionsf() and g()- Recall that (9) is fixed in place and that 6) is shifted backwards by ¢ For a given valle of f, the discontinuities in f+ 1) and a) define a set of rectangular strips, as was illustrated in Fig. 3. ach rectangular stip as fat the top and g atthe boom of vive vera. The side ofa stip ate determined by discontinuities in f and g. For the purposes of the algorithm, we separate the sis into four groups based on the discontinuities athe sides of the sis: ‘Re for those with Fon bath sides; Ry fr those with g 00 both sides; Ry for those with Fon the Let and gon the right; and ZR for those with g an the lft and fon the sigh. The sets Ry and Ry are parculrl important, a8 these are the strips whose widths change as f changes (as fis shifted) Thus, these strips fect the slope of M0) ‘We keep track of two quantiies: Hy and Hy Hy isthe sum ofthe squires of all the heights ofall the sips in Ry, and Hy fs the sum of the squares ofthe heights ofall the strips in Ry ‘The algorithm is based on the abservation tht for values f ¢ between two critical events the slope of 1,0) is Hig ~ Hy. This follows from the fact tha, fis Shifted backwards by 1 Ry is the set of all strips that increase in width by t and Ry is the set ofall stripe that decrease in width by The widths of the Ry and Ri, strips remain unchanged. ‘Consider what happens at one of the enitical events, where the change i no longer simply linear. We claim thar quantities Hy, and Hig can be easly updated at these points. To see this rote that, at critical even, 2 gFtype stip disappears (is width {oes to 20) and a new fe-iype sp appears (see Fig. 3) At the Sametime, the right boundary of the agjacent strip to the Tet is Converted from g of, an the lelt boundary af the adjacent strip to the right is converted from fo g. To update Hy and Hy we ‘ed to know jst the values of fan g around the ertical even, “This gives os the following algorithm. 1) nize + Given the piecewiseconstat functions fandg, dter- mine the critical event: the shifts off by 1 sch that {discontinuity inf coincides with «discontinuity in Sort these ertel evens by how far f must be shifted foreach evento occu. Let c,c,+-=e. be the ordered list of shits for the critical events c= 0. + Calculate (0,0), This involves summing the contibu- lions of etch of m+ srps and takes linear time, + Determine intial valves for My and Hy. 2) For + Determine the value of Ne + Update My and My: ) = (Hy Hype 6.) + Me-1.0) “The algorithm kes advantage of the fact that M(,0) is piecewise-linear as a function of f thus, the entire function fan be determined once we know an inital value and the slope for each piece. It is easy to see that the time for Inialzation is dominated by the time it takes to sort the iitical events: O(eloge), where e is the number of eritcal events, or Ofmnlogmn) where m and m are the sizes of the two polygons. ‘The updates required for the reminder of the algorithm take a total of O(e), or Own) time. a In practice, it might be useful to sealeuate (0) periodical from Scratch io avoid eros that could accumulate If thsi done ‘very O22) steps then the time bound forthe entire algorithm remains O(e loge). TV. Exaunes In this section we illustrate some of the qualitative aspects of the distance function (A, B) by comparing some simple polygons using the algorithm described in the previous section. In ‘Maition to providing a distance dA, B), between two polygonal shapes, the method gives the felatve orientation @ and the Corresponding reference points of the two polygons for which this distance i attained “The first example compares two simple polygons that are very similar in shape, but which are a different orientations (Gee Fig. 4). The value of di(4, B) is 0.144 which is attained fata ration of 180 degrees and withthe upper lft verex of the fist polygon matched with the lower fight vertex of the Second one. (Distances less than about 0.5 seem to correspond to polygons that a person would fate as resembling each oer: pais of polygons that are very diferent can have arbitrarily high sistance) To illustrate how the distance fusion can be used to compare model with several efferent instances, we consider the eight Shapes illustrated in Fig. S(a) and 5(b). In Fig. 5() the shapes tre ordered by ther distance from the square; in 5(b) the same Shapes are ordeted by thie distance from the triangle. The order ofthe shapes corresponds remarkably well fo our intuitive idea of shape-revemblance. The match tothe cutoff triangles suggests ‘thatthe metric is useful for matching partially occluded object, 15 long as the overall shape ofthe object does not change too racial ‘Our metic sIs0 provides « qualitatively good estimate of a rmatch when one polygon is an instance of another, but with, Some perturbation of its boundary. A simple example is given by the cutol ingle in Fig. 5. Another example is given in Fig. 6, where we compare 4 model rectangle against another rectangle with 2 notch removed. The distance is 0.327 with a felative orientation of 179 degrees. ‘An extzeme case of matching distorted polygons is shown in Fig. 7, where a triangle is compared with 2 somewhat tiangul shape. In this case the distance is 0834, and the orientation Uiference is 16 degres. Note however, hut, as mentioned in the introdvtion (oe Fig. 2), such perturbations must occur relatively uniformly along the perimeter ofthe polygon forthe match to bbe reasonable. (A smoothing technique is Iikely to alleviate the problem of nonuniform perturbations.) V. Sunny avo Discussion ‘We have suggested using the; metric on the turning functions of polygons as a way 1o implement the intuitive notion of Shape-resemblance. This method for comparing shapes has the following advantages + It is a metric on polygonal shapes. Ti compares shape alone; i i invariant under translation, rotation, and change-o- sale, + Tes reasonably easy to compe, aking time O(onn log min) to compare an m vertex polygon against ann vertex polygon. + Finally it corresponds well 10 intuitive notions of shape resemblance | a [> J IDE Yb SE Pepe Ode eS. Comparing sever polygons Fig. 6 A rectangle with oh removed /\ laa Fig. 7. Matching angle iy iy shape, In edition, this metic works for nonconvex as well a8 convex polygons, and even works for polygonal thapes that ae not simple. ‘A number of other authors have considered the problem of ‘etermining the extent to which one shape resembles ance (eas (11 3), [5], (8), [10], [11). In contrast 10 our distance function, these methods either ae not metrics, do not compare shapes independent of positon, orientation and sale or are ot ‘ficient to compute, The mos similar method to ours is that ‘of Schwarta and Share (11, [12] Huttenlocker and Kedem [4], have recenly developed a metic for comparing polygonal shapes independenly ofan affine transformation, thas extending out resus toa more geeral class of shape tansformations. Tacit Imetbod computes & distance based on the Hausdorff metic, and alio runs in time O(mnog mn) Like the method of [11] (12) our method is actualy based on 4 convoltion. Recall hat the major portion of our algorithm is devoted to minimizing A(t#) = J, ((s-+#) ~ ols) +9)ds. ‘When this formula is multiplied ou, all the tems depend con f alone of g aloe, except for the convolution term fi (a+ als) ds. Ifa ate piecewise-constant with m and 1 dscontinuies, respectively, then each erm ean be eaeulted in either O¢m) or Oa) time’ except forthe convolution term, whic sems to require (mm log min) time OF course the fast Fourier transform (FFT) can be used to compute a convolution in (Elo) time, but his requires evenly spaced sample points foreach off and For our problem, the discontinue ate not necessarily evenly spaced, so the FFT cannot be used unless we are willing to approximate ovr functions f and g. A good ‘approximation may require mre than mn poins. (Schwartz and Share (11] avoid these discontinuities by rotating the turing functions. This makes it possible to use the FFT, although it rests their method 19 Convex polygons) In any case, the development of a fast method for convoltions using unevenly spaced sample points would lad to improvements in the time ‘bound for our technigue ‘We vsed the; metric, bu similar techniques can be used to develop polygon-resemblance metrics that are based on ifleret funetonsspace metrics, Unfortunately, not ll soch metrics have {y's advantages of being reasonably easy o compute and match ing our inttive idea of shape resemblance. For instance, iis also possible to compute the L; metic on two (9) functions Using an algorithm similar to that im Section I In the case af the L, metic, however, the value of @ is ot given direcly for each value of a ti forthe Lz metic. This foreach of ‘the mi rita! evens, the optimal value of @ must be computed explicitly. Using data stuctre similar vo tha in Setion 1 the ‘overall compulation ean be done in ime O(log n), a opposed {0 O(n"logn) forthe Ly metric The Ly metic has an additional drawback: the optimal match will our when one side of polygon Ais atthe same orientation 4 some side of polygon B. This is beewse D:"P(t0) is @ Piecewise linear i both ¢ and 8,30 the minimum occurs ata {tical even in fas well a at aestcal event inf. In conta, the Ly metric finds the optimal oricatation (in a least squares Sense) without equiring any two edges fo be identically oriented Examine Fig 810 sce why requiring identical orientations can be undesirable; for the £4 metic the best match occurs at an orientation vifference of 76 degres, binging two edges into alignment. This would rotate the two figures s0 that they approximately form a star, «bad match. In contrast, forthe ‘etic the best match sat an orientation ference of 7 degrees, ‘which agecs gute wel with ou intuitive sense of the best match, may be posible to apply our methods to problems involving parially occluded objects, that i objects for which the entire ‘model is known, but for which only potion ofthe Boundary sppear inthe image, Our technique as presented ere has ot been designed to work with such objects, although, ax shown by some of ourexamples, i seems to give intutivey cores answers A Fig. 8A sh for whch the L mee does pony bute Ls does we. ‘when objects are not severely occhided, The combination of ‘occluded objects and our desire to make ovr metric independent Gt change-ofscale causes some dificult, We were able to fontol change-of-scle problems by normalizing Out polygons to make all perimeters have length one. If portions ofa boundary fate unknown then it fs unclear how this normalization sould be done Ifthe seal ofthe image is Known, then partly occluded ‘objects should not present any severe difficulties. ‘Our results can be generalized to include cases in which the shapes Avand B have some or all of their Boundary represented as etelar ars. IFA includes some circular ates on its Boundary. then the turning function © (s) is piecewise linear instead of plecewise-oonstant. As befor, to compare shapes A and B we feed to minimize worm f este) - 8th “The derivation of 4") docs not change at all, beeause Ni) is stl 8 quadratic function of @ But he shapes ince circular arc, then h can be piecewise-cbic at function of # fr fixed 8 Gnstead of simply picewin- near. Thu, the minimum value OU f(y8) doesnot necessarily occur aa eal event and wore {information ie needed in order to determine the behavior of beeen crcl evens However, since h(t 6"() is piewewie-ubic as a function of f we can determine the Behavior of between cfd evens if we have enough dats points for AIF we eallst two addtional (1, 2(¢.6°(1)) pais betwen each adjacent psi of efile evens en we can detrmine the coefficients of Nea"(0) (e+ os at + ay) between ciel eve’) and compute the minimum aaiycaly. Ths te basi approach ‘calling (8) in Ot +) tier each of On) “ales sl valid and there an algorithm that rns in tie ‘Otman{m-+n)). The Ofma tog mn) ine algorithm ean ao be isnealzed 0 work with cicalar ac, nd th esoling sgt fa the same asymplode tine bound. The numberof updates needa andthe resulting accumulation of roxndoll err may De problematic in pace. Acxsowesnanenr ‘We woul ike to thank K. Zikan for suggesting the Ly norm lover the Ly norm. Resowsncts, 10) D. Ais and 1G, “A combnstori gpanh polygon ry IE Irani zo vl 31983 (21 BH Big an Ca" Cnr Swan [3] PCox, H. Maite, Mi. Minoux, and C. Ribcito, “Optimal tang of over palo Pate Recoguton Lt YO 9, pears (4) DLP, Hutenlocer and. K. Kedem, “Compating the, minim Hatsdof diane fr pon ses under Uansition~ in Proc. ACM ‘S\mp Computational Gooner, 1990, pp. 380-30 151.) Hone tnd Tan, he’ smarberecen shapes under tine tsformation” in Proc Second Int Con. Computer sion ‘Washingt, DC: IEEE Compat Soc. Pres TH, pp. 80395. (6 ion Won Ae oer led ouch ing method ung form,” it Proc Nin I. Con, Patern econ Roe, ely Ny. 18, 1988 17] Br kde, The problem of et Shape descrip.” ip Proc 7 Pant Coy Computer vison. Washingion. DC IEEE Compt. Sic Pres, 1087 pp. 2606 [8] O'Rourke and Wishegton, “Curve sii vi ia Computaronal Geometry G- Tous EA. Am Retin. Nort Holla, 1985, pp. 33-318 i Royden Real Analv,” Nev ork. Macmillan, 1968 LLG, ‘Stipub and RCM, Haale, "Orginiation of ehtionat ‘ods for acene amavis” FEEE rans Pater Anal Machine Tet, vol PAMM-2 0,9 595-600, 1982 JT eSchwase end M. Share "Some smh o obo vision” New ove Une Coram na. Math. 'Sc, Teh. Rep. 119, Robotics Rep. 28 Ape 18 ep Pidetscation of sects in two andthe dimensions by atchng nosy chutes curves Ine J Robots Rey 0.8 fo 2 pp Bae, ST 1 Sill eel Game, New Yn My, 136 #i lic, by carve thing n Proc IEEE Mashop Com pur Vision, Mist Beat PL: Now. 30-Dec-1987 cm, The ether M, Arkin rceved the BS. depee in hnuberats ome Aviv University i 1981, nd he MS. and PHD, degree im operation ‘exec fom Stanford Univesity fn P83 and 1986 respecte. ‘nce grdunion in 1886, she fas been a visting Resistant Professor inthe Scho! of (Operon Reseuch snd Indu Eines SP Coral Univers, where she conducts ee Starcpeaph thor, competstonal peomesy, Shed. tnd optimization, 1, Paul Chew sccived the BSc. snd MS. (gees ia mathematics andthe PRD tepiee Iompute scence fom Pusdae University To7a BPS an 196, reaped He wars fealty member st Darth Col lege uni 1948 when be began his caren po ‘Son is a Senor Research Awe the Deparment of Computer Scence at Cortll Unveils rsa interests acolo Puna geomery espn and anal of feet and mes geeraton Daniel P. Huttenlocher received the BS, de ‘ee tom the Univer of Michigan i I980 {ng the MS.and Ph, degrees om the Mass: emus insta of Technlogy 1988 and To respectively THE is Carell an Assistant Professor inthe Deputoeat of Computer Science at Covell pverty, ands member ofthe Reseuch Stl AC the Xeon Palo Alto Reserch Center Hie ‘serch itis re in the ste of computer ‘sion, computational geome speech C08. ‘hoo, and seal lignce Klara Kedem ceived he BA, dec in mits and pio InviSri te MSc" degre in compte since, n 1082 ander "ae ‘perisin of professor A Puy fd te PR. deen 1989 inet emrgangt tea MO ih Ucn sclmute sate in 999 th held Viking Sees postion at Cornell Unvesy.Cuemty she i ntact postion ae Av {ry ig sch ptt geome al apleaons {m'robtie an compar vse Joseph .B, Mitchell sesved the BS. degree inapplid mitemates and physics and ie MS. Segre in matvemates fom Caoepe Mellon Univesity, Pasutph. PA oth In 198) and the PhD, degree in operations research fom Stef Unneriy in Tob ‘During Bs tine a! Sunn, be worked st tne Hughes Arica Ineligenss Resch Cee ter. Sites gration. he bss een an Amis é Profesor of Operations Reseteh and Inti Enginceing a Cornell Univer is primary research inert at ia the els of computthna eam, Seen of Hgts, compatr vio, td opeaton ‘Dr. Michell a member ofthe Assocation for Comptig, Ma hier, the Operations Research Society of Amica, andthe IEEE Computer Sot

Anda mungkin juga menyukai