This action might not be possible to undo. Are you sure you want to continue?
Robotics Research Ibchnical Report
CustomMade Pyramids
by
Shmuel Peleg t Orna Federbusch
Robert A.
t
Hummel
%
Technical Report No. 255 Robotics Report No. 85
November, 1986
\,
0)
New York University
Cc o
S

I
nstitute of
Mathematical Sciences
pij
O
Computer Science Division 25 Mercer Street New York, N.Y 00
1
1
1
^^
.„.A
„..iS
_,£i
^.
.:^.:^:^
j:l
:^^.^^.^.,^*;j*:^: f*"" j^^^f^^
<
K. ^^.:^^»^^':[
Support for Dr. New York 10012 USA hummel® nyu. ONR . t Hummel i Technical Report No. Israel peleg@hujics. 255 Robotics Report No. 85 November.csnet tCourant Institute of Mathematical Sciences University New York New 251 Mercer Street York.arpa This research has been supported by a grant from the Israel Academy of Sciences.bitnet or peleg® israel. 1986 fDepartment of Computer Science The Hebrew University of Jerusalem 91904 Jerusalem.CustomMade Pyramids by Shmuel Peleg t Orna Federbusch Robert A. Peleg's visit to New York University was provided by Grant N0001485K0077.
.
In the Burt form of the Gaussian pyramid [2]. pixel it is on every row and offsetting the samples by possible to achieve an effective sampling ratio of . while higher Pyramids are data structures used levels of the pyramid are used for data stored at coarser levels of resolution. the reduction factors can be made adaptive to region properties. By subsampling every other one pixel on successive rows. The reduction factors can be different in each dimension. Pyramids Pyramid data structures have proven useful in the development of image processing methods dealing with image features at varying scales of resolution. the reduction from one level to the next is accomplished by blurring the lower level (by means of convolution with a nonnegative kernel) followed by a subsampling of every other pixel on every other row. Further. The user can thus specify a sequence of decreasing rectangular image sizes. Typical image pyramids are formed from rectangular grid arrays whose sides have lengths that are powers of two in extent. and construct pyramids conforming to those sizes. in the Gaussian pyramid data structure. and differ between levels. to adapt to a given appHcation. so that in the subsampHng operation left and right edge pixels are included on every second row. and we say that the size ratio between levels is two in each dimension. For example. Thus the base level might be 512 by 512. each successive level is obtained by local averaging and subsampling of the immediately lower level in the pyramid.CustomMade Pyramids Shmuel Peleg Orna Federbusch Robert Hummel Abstract to store and process images at multiple levels of resolution. enabling smooth regions to be reduced more than "busy" regions. the size reduction in each dimension between levels of the pyramid is a constant factor of two. 1. each level has size of the form 2" + l by 2"+ 1. the sampling rate is still two. This paper describes a scheme that permits construction of pyramids with arbitrary size reductions between levels. the next level up would then be 256 by 256. In nearly all pyramid implementations to date. A survey on pyramid structures may be found in [1]. and each successive level reduces each dimension by a factor of two. and that both the top and bottom rows are sampled. In the Gaussian pyramid data structure. Nonetheless. The bottom level of a pyramid is used to represent data at a fine level of resolution.
.
Further. treat < N and both cases . In this paper. the reduced reduction factor between levels results in a better chance of capturing a salient feature at an appropriate scale of resolution. taking contributions of fixed amounts. the reduction does not have to be uniform over the whole image. we describe a method for constructing at level. OneDimensional Resampling We els first to a vector (wo." and the new sample points that form the resampled data as the "consumers. corresponding to the sample point locations in the domain. such that the total of all contributions by a given producer is the total amount produced at that site. and then sampling the columns of the result. consider the original sample pomts as "producers. The consumers obtain their products from a linear sum of the producers. As an introduction to the resampling methods to be defined in later sections. The error is minimized when the consumers use the closest suppliers to compute their values. at the cost of doubling the number of levels. but we will not pursue the variational derivation any further.1. as is the consumption level of each consumer." The producers and consumers have associated positions. every other level is in located on a 45° grid. Uniform Resampling We use the formula r(<+i)/pii j= Li/pJ where Page 2 . and care must be exercised local operators when designing kernels and essence other Nonetheless.y^M\) • • • • • • ^ M ^ N. .5]. The central property of the sampling fully to the output method is that each pixel contributes samples. The amount "produced" by each producer is prespecified. and can vary based on local image properties. will then present a formulation of the sampling idea that permits the construction of pyramids using arbitrary placements of pixels (such as hexagonal to rectangular grids We grids). The transformation can be applied by first sampling the rows. M consider the problem of resampling a vector (vq. pyramids of arbitrary size needed.CustomMade Pyramids \/2 in each dimension [3]. In this pyramid scheme. and not only by fixed reduced as each be ratios. thereby minimizing sampling artifacts. The methods described below have certain optimality properties with respect to these measures. The "error" of a sampling can be defined as the sum over all producerconsumer pairs of the distance between the corresponding points weighted by the contribution of the producer to the consumer. 2. related to antialiasing. The is a spatial will first We describe the resampling idea in one dimension. so that the resolution can basic step in the construction of the proposed pyramid scheme resampling technique used in graphics [4.vati) of A^ pixWe of pixels. 2. on these levels.
( ( pO + 1) (z  /) } } if if U/pJ z7p ^ . .1. "^ 1=0 rij = = p. / + 1]. ^ r^ 1. p'(y + l)].. we can regard the output range [0.. so that p is small.p. [/.^ < A^. For the situation with N ~ 4 and (wo. Uniform resampling weights for a 4vector to a 3vector. with the exception of the two extreme n/s.A/] split up into M subinterM— 1. with j = 0. 1 wi = —vi + —V2 W2 = 7V2 + as "TV3 Figure la. More generally. • • • . p. Figure lb shows the overlapping intervals.Peleg. . = VQ Vj V2 V3 Figure la. to resample (vo. 1.V2. /V — 1. Then all for j in the appropriate range will have the value p.vi.1. p(y l)] that intersects with the interval vals • • .. in shown how the coefficients can be computed from Properties of this sampling method include: Nl • • The total contribution of any given Vj to all w's is p.e. The same range is also split into A^ subin[/. / + !].V3) with Wo = 3 jvo + —VI .is one. p.^2) we have the formulas M = 3. tervals [pj. Page 3 . The coefficient rij is the total interval + length of the portion of the [p. Ml The sum of all contributions to any given w. Federbusch. and Hummel nj = ^ min { 1.) < < r(/ + l)/pl and P = assume that A/ To nearly interpret this formula. with / = 0. < i'P min{ + 1) . i.e.. first r. w 1. i.
CustomMade Pyramids HO .
Peleg. Pyramid representations of the weighted sampling with weights as in the text. and Hummel vo VI Figure 2a. MO . Federbusch.
els 9 1 1 \ / I \ I ^ 1 2' / 3 1 6/^6 \ / / 13 I I / I \ I \ x \ I I / / I I I 1/ \1 <5 b Figure comparison of a 432 one dimensional pyramid with a 42 pyramid leftmost pyramid shows the weights in a 432 structure. and that the supports of the linear weighting functions can overlap in lower levels.1 or 1. although in the onedimensional version of his scheme (using fivetap filters)"/ adjacent pixels in one level share supports from three pixels in the immediately preceding level (see Figure 4). Successive lev have larger overlaps in the base level. Burt's overlapping pyramids [2] also exhibit overlapping supports. resampling using either the uniform resampling or weighted resampling of Section 1.1 or 1.e. depending on the structure ofihe intermediate levels. j ^ = r^ — 1. followed by a resampling along the An input image of size M by N can resulting rows. However. The resuh is the same if the columns are first resampled. and are not adjustable in the same way that the taps in the Burt pyramid can be modified.2 yields weights determined by the sizes of the levels. The Page 6 . applying first a resampling of the rows (which are ^/vectors) to have lengths L. we most one pixel. Pyramid Construction be resampled to any desired output size K by L using the uniform resampling method of Section 1. The rightmost pyramid gives the weights for a uniform 42 resampling.CustomMade Pyramids M1 • The sum of all contributions to any given w. and then resampling the resulting L columns (each of which are Mvectors) to have lengths K.1. In the immediately preceding level in a onedimensional pyramid. Note that the values in higher levels are weighted averages of bottomlevel pixels. A built by specifying the desired present an example of a threelevel onedimensional pyramid. while the middle pyramid shows the effective weights from the bottom level to the top. However. 3.. the supports from lower levels can have arbitrarily large overlaps. support functions for adjacent pixels can overlap in at In Figure 3. A structure. sizes of all complete pyramid structure can be levels above the base level.2. is one. 3. with resampling as given in Section 1. i.
however. that uniform resampling will be used.. for the time being.a . proceeds as follows. Recursively.(x)'a as in Jifcction 2. for z < €. We construct the Laplacian pyramid as a sequence of images L.(x) all have unit mass. or for expanding a level to a larger image. + i: Page 7 . and Hummel Figure 4.. + i by A/... . where 7Z is the resampling operator. we instead simply assume that the interpolation kernels <}>. Federbusch. + i is obtained from G. resampling is still done by the formula w/ = but now the r^y's are given by fij = J^i(x)^j(x)dx. • • • . we mean that 'i 1=0 S <t>/U) ^ 1 and f}\ii(x)dx = 1.and an expansion resampling of G. the total of all contributions to a given pixel from given lower level will be unity.of size TV. of decreasing sizes (A^.by resampling from an A^.. pixels at some The pyramid resampling scheme can be used size levels (as in a for building successively smaller Gaussian pyramid). 1. all In all cases. = 0.Peleg. € of a • • • Gi + l = T^(Ni X Mi) . for building analogs to the Laplacian By combining structure.by M.•) f'or. 1.. Let Gq be the Nq by Mq initial image. I €. By this.>uiiition of unity of the domain. M.. the resamplings may be used a pyramid data The • construction. In this case.image to an /V. analogous to Burt's formulation. '^rijVj. a subsampling operation for contraction with expansion operations.) for levels pyramid structure. by M. and that the sampling kernels v]/.by M.{Ni+i X Mi+i) (Gi) . Using the interpolation formula and sampling method formulation of resampling.(x)'s and i}j.•(. Rather than defining (t>. + i image: specify • • We sequence /. I = 0.. We assume. G.is defined as the difference between G.. for 1 = 0.2. Then L. The Burt 5tap pyramid. we can gencraiizc the uniform and weighted resampling methods to permit larger overlaps in the support functions of the resampling kernels.
at the top level. i).•. level L.L. • • • . Recall that the resamthe data (vq.'s using the (obvious) formulas As Ge = Li G. Go can be recon with the Burt Laplacian pyramid. . . Thus uniform resampling for contraction is paired with uniform resampling for expansion. the original image structed exactly from the L.. 2). Figure 5 shows such a custombuih Laplacian pyramid structure. . . Mvector (wq. For example. It is then desirable to invert (in some sense) the weighted resampling that was used in the contraction from level G.an iVvector (vq'. 1._i + TZ(Mi X Mi) ~ (Nii X Mii) (G) • Each represents an approximate bandpass filtered version of the original image... 1. using )the levels. • • • . • • • . + i for the expansion operation in the construction of L. We now define the "backward weights" P... where the widths of the bands are determined by the choices of the relative sizes of the levels. represents the sum of proportions of v/s that contributed to w.Pw . Accordingly. we showed Page 8 . in Figure 2 the weighted sampling of a 4weights (1.. so that the weighted backward resampling becomes As another example.wjizi). and have value p"^ = N/M. Assume (po. uniform resampling of an A^vector to an Mvector corresponds to weighted sampling where the weights are all equal to p = M/N. A weighted expanused in the reconstruction of G. the vni') which expansion using weights (Po.to G. j= i^j Ml It is easy to see that 2 i=0 • • Pi • ~ ^' ^° ^^^^ ^^^ P''^ ^^^ serve as weights for resampling to • the Each p. i = 0. The resulting backward weights vector to a 3vector with are (1.\i'Nl) has been resampled using weights to obtain the A/vector (wq. we define L( = Gf. What weights should be used? . The backward weights are then also all equal. skewsense that there is no horizontal in the . Finally.vn best approximates (vo.VAr_i'). Suppose that we wish to use weighted resampling in the construction of the Laplacian pyramid.l = L.vni) • • • pling is linear: Nl_ j = The matrix according to coefficients r^ represent the contribution fractions of coefficient vj to coefficient w. The same weighted expansion should then be L^. • • that . M—l.CustomMade Pyramids ^i — ^i ~ '^(Ni + i X Mi + i) ~(Ni X Mi) (G.l) leads to a vector (vq'.+ 1 sion resampling should be used (even. . 1/2). 1/2. + i) . ^ .wmi) • • • • • • • • • • ing. though it is not absolutely required for exact reconstruction) so that the levels Lf m^kHauftheif approximate bandpass characteristics.
Federbusch. and Hammel Figare 5a Page 9 .Peleg.
CustomMade Pyramids Figure 5b Figure 5. The absolute value of the data is displayed. wo vi' V2' vo VI V2 = Wl = H'2 + Vi 2 V3' = W2 V2 + V3 Page 10 . Levels are enlarged to full size for visibility. with larger values shown more darkly. Figure 5a show the Gaussian pyramid. A custombuilt twodimensional Laplacian pyramid with level sizes as marked. while the Figure 5b contains the levels of the Laplacian pyramid.
Po = 1 . as introduced in has the advantage that regions with "interesting" activity can be resampled more finely than "uninteresting" regions. tessellation grids. we will use a smoothed absolute value of the Laplacian of the data to measure "busyness. The total effect of the contraction followed by the expansion is that the higher weighted nodes were exactly reconstructed. as depicted 4. and also discuss extensions to two dimensions and irregular Section 1. Adaptive Resampling Weighted resampling. 4.Peleg. In this Section.2. we suggest an interest operator for obtaining the resampling weights. Figure 6a. Federbusch. Onedimensional Adaptive Pyramid We has been observed that suggest an interest operator based on the local "busyness" of the data. Backward weights associated with the weighted resampling of Figure 2.1. This can be accomplished (in one dimension) by giving interesting pixels larger weights. Here. as in Figure 7. It in human perception a line with higher "busyness" seems longer than a straight line segmsnt [6]. and Hummel by Figures 6a and 6b." A similar operator has been suggested for representation of intensity information by retinal receptive fields [7]. while the lower weighted nodes were blurred.
N2 + ^bt^Ni NbT. The ^i.+i i . One way to accomplish the smoothing is to define an iterative weighted smoothing as follows: and then smooth the • • • results = boj = lj. Specifically. by any reasonable local averaging operator. given a signal t'c. An illusory distortion in which the divided space appears longer than the undivided space (OppelDundt illusion).Nl = . The sizes of the levels are prespecified. • • • . I = 1.'s.<A^2.i  2v. The signal data is given. + v. / = 0. l<. and normalize to obtain the values to include the endpoints / — x.o 4' "'' T^' 1 ' bt+i. and weights are obtained from the busyness measures of that data.. and recursively set 4 fcf+1. v. ^Nl bo.N l.CustomMade Pyramids Figure 7.0  T^'. the absolute Lapla cian values: ^i = I v. .i \^i= TTTi j = for some fixed integer 7 > representing the amount of smoothing.'s are used as weights for resampling the adaptive custommade onedimensional Laplacian pyramid can therefore be easily constructed.^'l. we compute A^2.va'i.Ni Finally. * * " . An Page 12 . extrapolate and i — N—l. set = ^bt.
j) in the A^ image. This must be done using the backward weights. Then the ' row weights. Suppose that we wish to resample to an N by M' image.Peleg. and the are used for resampling the columns. ' = Nl M\ k=Q 7=0 (xo. Federbusch. Twodimensional Adaptive P^^ramids The twodimensional case involves technical difficulties not present in the onedimensional case. One solution is row or column Specifically. N'^bij V. continuing in this fashion yields a Gaussian pyramid with adaptive weights. To obtain the Laplacian pyramid.2. whereas in two dimensions it is not clear how to chan^^e the size of a rectangular region in relation to a weight and still have a complete decomposition of the domain. by M we compute a "busyness measure" bij at each pixel (i.\lm]) are used for resampling every row. each level above the base must be expanded to the size of the previous level. used to resample an Mvector to an M' vector. the row weights will be the average row vector of a busyness matrix. nearby pixels may give principle contributions to distant pixels in the next level. then the image will lose its structure in the sense that neither connectivity or convexity of shapes will be preserved. is to resample "n each dimension separately. 4. i^ — • • The same weights weights (vo.vn\) Page 13 . The resulting expanded levels are differenced with the Gaussian pyramid data at that level to obtain the Laplacian data. Recomputing new busyness measures gives new weights that can be used for the next level resampling. However. and Hummel Resampling using these weights results in the data immediately above the base level. are given by Nl M'^bij 1 = y^j 1=0 t=0 to and the column weights. and the column weights will be the average column vector of the same matrix. Reconstruction from the Laplacian pyramid structure is possible. weights can be converted to interval lengths to decompose the domain. • • • . Since the weights on one row may be independent and different from the weights on a neighboring row. used are given by resample each column yVvector to an //'vector. In this case. as described in Section 3. . The problems avizt due to the fact that in one dimension. we suggest a scheme where all rows and all columns use the same weights. This means that the complete adaptive Laplacian pyramid data structure consists of the Laplacian data at each level together with the backward weights needed for the expansion resampling at each level above the base. Since this behavior is typically unacceptable. if each processed independently. but requires that expansion resampling using the appropriate backward weights at each level be used.
:i row and column weights are both (1/2. Hexagonal vo Vl V2 V3 Figure 8a. 5. 1). a possible method for is to smooth to a 2 As a trivial example. Once again. 2/3. The nonbusy quadrants (the upper left and lower right portions) of the 4 by 4 image can not be contracted without distorting shapes in the remaining portions of the image. Irregular Tessellations The nonadaptive versions of the resampling methods given in previous secgrids tions easily extend to grid arrays other than rectangular lattices. 1/2). Figure 9 shows an adaptive custommade pyramid of an image. and the column weights are 2/3. 1/2.CustomMade Pyramids computing "busyness values" absolute values of the Laplacian of the image data. and column resampling is as shown in Figure 8b. if the busyness matrix has the form 2 2 ' 2 2 • '"'""' ^v ' ' 2 2 2 2 In this case. Row resampling using adaptive weights of the simple example. the row weights are given by (0. and that the other pix A contrasting example arises. 1. Page 14 . suppose that we wish to resample a 4 by 4 image by 2 image. 1/2. 0. 2 by 2 image. The row resampling will then work as shown in Figure 8a. 2/3). and the resampling for both the rows and columns will be uniform. Note that the result is that the 3 by the a 2 bottom right subimage will be resampled to els will be ignored. the . Note how the interesting regions in the image become stretched into larger windows at the higher levels. and that the busyness matrix has form 2 2 2 2 2 2 Then (0.
Figure 9a Page 15 .Peleg. Federbusch. Column resampling using adaptive weights of the simple example. and Hummel Figure 8b.
There are hexagonal grids where the cell sizes expand with increasing distance from a central cell We [8]. might also imagine a random placement of sample points. ^ •ti^^ Figure 9b Figure shows the pyramid. and the Figure 9b shows the Laplacian adaptive pyramid. and such grids may have a basis in human retinal cell distributions [9]. We will first discuss the general case. six A level adaptive levels of the are often suggested as a useful arrangement for image processing. with local density of sampling points prescribed by some rule (perhaps adaptive to the image data).. 9. Figure 9a Gaussian adaptive pyramid.CastomMade Pyramids V /K ' ' » . and then focus on a case where sample point locations are described on a polar grid. with level sizes as marked. Page 16 .
represents its corresponding cell C. functions of the cells and sampling functions that are characteristic normalized to have unit mass. should prove adequate for effective resampling.}/lo^. of course. Page 17 .. but the simplest such kernels. More general interpolation and sampling kernels can be envisioned. the polar sampling. then data defined on the C grid to the I>grid as follows.Peleg.1. the locations of the sampling points are associated with either a regular tessellation or irregular tessellation (a decomposition) of the image domain into disjoint cells. Placement of the circles with radii increasing exponentially [11] and also simply increasing linearly [12] are also known (see Figure 10). • • • . . and the union of all such cells covers the domain. where resolution falls off with the distance (eccentricity) from the fovea. otherwise and i^i{x.2. Given one such decomposition of the domain. each Vj representing a value define resampled data (wq.) = JJ^dxdy . Each sample v. We . the is We now intersections important for sample points are located on the Some models of human retinal receptor distribution describe samples points in this way. If a tessellation is not provided as part of the sampling locations.y) that are characteristic func tions of the corresponding cells Cj. interpolation functions ^j(x. C j. then one can be provided by constructing the Voronoi decomposition associated with the points [10].y)€Vi otherwise Area. area measured by integration (2?.y) = I ^ 1 (x.y)^Cj . Federbusch.w»y_i) by • • • • • Nl where 'y "" SI^i(^'y^^j(^'y^'^"'^y in Here we have mind P.y) = I ^ is l/area(r. That is. if y\ijix. between a set of rays and a set of concentric circles. say {D. we can define a resampling of We on cell are given an A^vector (vq.}fLV» and another (perhaps coarser) decomposition. where A and h are constants [9]. (x.. which leads to placement of grid points on concentric circles whose radii increase as \og(A+kh).2.vv_i) of data. say {C.) if . A: = l. which In polar sampling. and Hummel 5. snapshot visual perception. 5. Polar sampling discuss a special sampling. General case In all cases. Sometimes linear falloff of the sampling rate is assumed. the characteristic function of the cells as described here.
reduced. the points on each concentric circle can be resampled independently of the other circles. Summary Using the antialiasing computer graphics leads to an idea points. while on the right the growth of the radius is linear. alluded to in Section 1. in Section 4. has to be extended to resampling along a circle. As with twodimensional resampling. Further. we simply resample the grid points along each ray independently of the other rays. where sampling points are on the intersection points between lines and circles.2. or we can reduce the number of rays. a separable kind of adaptive resampling is possible for polar grid pyramids. because the polar grid coordinate locations can be decomposed into a cross product of circles and rays. Extending this idea to for onedimensional resampling of N points into two dimensions easily establishes a method for building pyramids with arbitrary in method common M Figure 10. Polar sampling. If both the number of concentric circles and the number of rays are to be reduced. the notion of onedimensional resampling of a line segment. and then by reducing the number of rays. we can then do the resampling first by reducing the number of circles. Indeed. the methods of Section 2 extend easily. similar to the adaptive tv. To do this. We using distance based on radian measure in place of onedimensional Euclidean distance. Page 18 . this process is unaffected if we exchange the order of resampling. The left sampling is with exponential radius growth. the optimality properties of the weights used for resampling. extend to polar resampling. Resampling a ray is straightforward using the methods of Section 2. When the number of rays is reduced. as described in Section 2. However.CustomMade Pyramids can consider resampling a polar grid by reducing the number of concentric When the concentric circles are circles.'o dimensional rectangular pyramids discussed 6.
In two dimensions. Structured Computer Vision. CMU Robotics Institute. B." Proceedings of the 20th Fundamentals of Computer Science Conference.D. Fant. D. Schwartz. and J. [12] pp. "Detection and recognition of simple forms. M. and T. and Robert A. Multiresolution Image Processing and Applications. "3D transformations of images in scanline order. Ames Research Center. Anstis. or placed on polar grid or hexagonal patterns." Computer Graphics (Proceedings SIGGRAPH 80) 14. Thesis (1982). The question We have also investigated the idea of adaptive resampling. Seeing is Deceiving. S. 1983).Peleg. References [I] Rosenfeld. K. Catmul. "A representation for visual information." IEEE 532 (1983). and allows for arbitrary nonlinear stretching along the line to be resampled. G. 7180 (1986)." IEEE Computer Graphics and Applications 16. [4] "A nonaliasing. Lawrence Erlbaum Associates (1978). Hummel. laplacian pyramid as a compact image code. (April. S. "The p. Ideally. on Communications." NASA Technical Memorandum 84353. and then the columns are resampled using a single set of weights.. [5] [6] Coren. and we compromised by permitting only a "separable stretching.. Finally. and Hummel We thus see that the use of pyramids with dimensions an unnecessary restriction on the construction of the as to what size levels are most appropriate is left unanswered. and A. P.. Academic Press. pp. Trans. S. sizes specified for each level. New York (1980). but our method does not easily extend to the case of adaptive placement of resample point locations in two dimensions. "Receptive fields and information. adaptive resampling would allow one to dynamically place additional points in regions with high busyness. is given by powers of two pyramids. Adelson. things are more complicated. [7] Zucker." Vision Research 14. note that the idea extends to irregular tessellations. 1827 (1979).. "Efficient computation of continuous skeletons. yielding the same stretching for all rows." Vision Research 20. 645669 A [10] (1980). and depends upon the application and empirical experiences. Page 19 . Steven of visual W.." Ph. This research has been supported by a grant from the Israel Academy of SciSupport for Dr. A. To appear. The idea is easy in one dimension. Kirkpatrick. where sample randomly placed. Smith. realtime spatial transform technique. pp. [3] Crowley. spatial Watson. J. Springer Verlag (1984)." where the rows are resampled using one set of weights. pp. R. A.. [2] Burt." Human Neurobiology. "Computational anatomy and functional architecture of striate cortex: spatial mapping approach to perceptual coding. pp. Eric. the reconstruction [8] [9] Tanimoto. E. (1985). [II] "A chart demonstrating variations in acuity with retinal position.. 279289 (1980). 589592 (1974). we points can be Acknowl edgements ences. We gave one particular busyness measure to use as a basis for deciding on the stretching. Federbusch. Girgus. Peleg's visit to New York University was provided by ONR Grant N0001485K0077. S.
CustomMade Pyramids Page 20 .
aOS^Qf^^DUE .
NYU COMPSCI TR255 c. . Shmuel Custommade pyramids.l Peleg.