3D Frequency-Domain Analysis of Non-Separable Reconstruction Schemes by using Direct Volume Rendering

Bal zs Cs bfalvi a e Budapest University of Technology and Economics Bal zs Domonkos a Mediso Medical Imaging Systems, Hungary

Linear box spline.

Trilinear B-spline.

Quintic box spline.

Tricubic B-spline.

Figure 1: Frequency responses of different reconstruction lters proposed for the BCC lattice.

Abstract

Introduction

Recently, it has been demonstrated that the optimality of the BodyCentered Cubic (BCC) lattice can be utilized also in practice by using either non-separable box-spline lters or tensor-product Bspline lters for reconstructing an originally continuous signal from its discrete BCC-sampled representation. In order to study the frequency-domain behavior of these lters, a 3D analysis of their frequency responses is required. In this paper, we show that direct volume rendering offers a natural tool for such a 3D analysis. As the frequency responses are analytically known, their characteristic isosurfaces can be rendered separately in the pass band and in the stop band. The visualization of the frequency responses conveys information not just on the absolute postaliasing and oversmoothing effects, but also on their direction dependence. In this sense, the frequency-domain behavior of the non-separable box splines and the tensor-product B-splines on the BCC lattice is evaluated for the rst time in this paper. Furthermore, we also analyze how the frequency responses are inuenced by a discrete preltering, which is necessary to fully exploit the approximation power of the higherorder box-spline and B-spline lters.

In volume-rendering applications, the input data is usually a discrete representation of a continuous phenomenon. In order to faithfully visualize the underlying continuous function, it has to be accurately reconstructed from its discrete samples. If the volumetric data is acquired by regular sampling on a specic lattice, the reconstruction can be easily performed by a convolution ltering. An appropriate choice of the lter kernel is of crucial importance as it directly determines the quality of the rendered images. Visual artifacts are often introduced because of the imperfect frequencydomain behavior of the applied lter [Marschner and Lobb 1994]. Theoretically, an ideal low-pass lter can perfectly reconstruct a band-limited signal if it is sampled above the Nyquist limit [Oppenheim and Schafer 1989]. The Fourier transform of such an ideal lter is a characteristic function that takes a value of one in the pass band, and a value of zero in the stop band. If the original signal is not band-limited or not sufciently sampled, which is often the case in practice, the ideal low-pass lter might result in ringing artifacts [Marschner and Lobb 1994]. This prealiasing effect is caused by the drastic cut-off in the frequency domain. Ringing can be avoided by using practical lters that ensure a smooth transition between the pass band and the stop band. However, the energy of the lter in the stop band still needs to be as low as possible to reduce postaliasing effects. For the Cartesian Cubic (CC) lattice, reconstruction lters are usually designed in 1D and extended to 3D by using a separable tensorproduct extension [Mitchell and Netravali 1988; M ller et al. 1998]. o Therefore, it is sufcient to analyze the frequency-domain behavior just in 1D based on frequency plots [Theul et al. 2000; Bentum et al. 1996; Cs bfalvi 2008b]. This approach, however, is not feasie ble for the optimal BCC lattice, since the advantageous properties

CR Categories: I.3.7 [Computer Graphics]: Three-Dimensional Graphics and Realism; I.4.5 [Image Processing and Computer Vision]: Reconstruction

Keywords: volume rendering, optimal regular volume sampling, reconstruction ltering

e-mail:

e-mail: cseb@iit.bme.hu balazs.domonkos@mediso.hu

of a 1D lter are not necessarily inherited in 3D if its separable extension is used on the non-separable BCC lattice. Furthermore, when a signal is sampled on a BCC lattice, its spectrum is replicated around the points of the dual Face-Centered Cubic (FCC) lattice [Entezari et al. 2004], which is not separable either. Therefore, a 3D frequency-domain analysis is really necessary to study how much the aliasing spectra contribute to the reconstructed signal if a given lter kernel is applied. The contributions of this paper are the following: We thoroughly analyze the frequency-domain behavior of box-spline and B-spline lters proposed for the BCC lattice. Previously, the frequency responses have been studied based on only 2D cross-sectional slices [Cs bfalvi 2010]. In cone trast, we render the frequency responses by direct volume rendering. Therefore, we can identify the preferred directions of a given lter based on the 3D distribution of its spectrum. We validate the results of our 3D frequency-domain analysis by rendering BCC-sampled synthetic and measured data sets. We demonstrate that the postaliasing effect is indeed minimal along the preferred directions identied in the rendered frequency responses. We conrm that on the BCC lattice, the box splines perform better in the pass band, while the B-splines ensure better stopband behavior. Therefore, the box splines can capture more high-frequency details, whereas the B-splines suppress the postaliasing effect more efciently. For the sake of fair comparison, we use recently proposed discrete prelters for quasiinterpolation [Cs bfalvi 2010] that fully exploit the approxie mation powers of the B-splines and the box splines as well.

lters have been applied for quasi-interpolation on the BCC lattice [Entezari et al. 2009; Cs bfalvi 2010]. From a practical e point of view, a separable BCC B-spline ltering can be implemented on the recent GPUs more efciently than a non-separable box-spline ltering of the same approximation order [Finkbeiner et al. 2009; Cs bfalvi 2010], since the B-splines can utilize the e hardware-accelerated trilinear texture fetching [Sigg and Hadwiger 2005]. Although these ltering schemes have already been compared both theoretically and practically [Finkbeiner et al. 2009; Cs bfalvi 2010], their 3D frequency-domain analysis has not been e published yet. According to our best knowledge, the frequency responses of box splines and B-splines applied on the BCC lattice are analyzed by direct volume rendering for the rst time in this paper.

Frequency Responses of Box Splines and B-Splines

In this section, we review the frequency responses of the nonseparable box splines and the tensor-product B-splines. As these reconstruction lters are usually applied on preltered data values [Finkbeiner et al. 2009; Cs bfalvi 2010], the Fourier transforms e of the discrete prelters are also discussed. The resultant impulse response of a preltered reconstruction is the convolution of the continuous reconstruction lter and the discrete prelter p. Con sequently, the resultant frequency response is ( ) p( ), where and p are the Fourier transforms of and p, respectively.

3.1

Box-Spline Reconstruction

Related Work

The optimal BCC lattice was rst used for volume rendering by Theul et al. [Theul et al. 2001]. However, they applied the splatting technique with spherical lters that resulted in rather blurry images. Mattausch [Mattausch 2003] studied several practical reconstruction schemes for the BCC lattice, but the results did not show unambiguously the theoretical benets of BCC sampling. For example, the sheared trilinear interpolation led to an anisotropic reconstruction. The rst reconstruction lters that are tailored to the geometry of the BCC lattice were proposed by Entezari et al. [Entezari et al. 2004; Entezari et al. 2008]. Their non-separable linear and quintic box splines provide the same approximation orders on the BCC lattice as the trilinear and tricubic B-splines on the CC lattice, respectively. Theoretically, the computational complexity of a box spline is lower than that of an equivalent B-spline, since its support is more compact and its total polynomial degree is lower. Nevertheless, this theoretical advantage cannot be exploited on the current GPUs [Finkbeiner et al. 2009], which are rather optimized for separable ltering. Another family of non-separable lters is represented by the BCC-splines [Cs bfalvi 2008a] that gene eralize the Hex-splines [Van De Ville et al. 2004] for the BCC lattice. The Hex-splines were originally proposed for the hexagonal lattice, which is optimal for sampling circularly band-limited 2D signals. The BCC-splines were evaluated by a discrete approximation and their analytical formulas are not known yet. Cs bfalvi rece ommended a preltered Gaussian reconstruction scheme [Cs bfalvi e 2005] adapting the principle of generalized interpolation [Blu et al. 1999] to the BCC lattice. This method was extended also to the B-spline family of lters [Cs bfalvi and Hadwiger 2006]. Both e the non-separable box-spline lters and the tensor-product B-spline

The linear box spline for the BCC lattice [Entezari et al. 2008] is dened by convolving a 1D box lter with itself along directions k that are dened as 1 1 -1 -1 1 -1 1 -1 1 . (1) = [ 1 , 2 , 3 , 4 ] = 2 -1 -1 1 1 Since the Fourier transform of a 1D box lter is a sinc function, and consecutive convolutions in the spatial domain correspond to consecutive multiplications in the frequency domain, the Fourier trans form M of the four-directional linear box spline M is expressed as a product of four sinc terms [Entezari et al. 2008]: M ( ) =
sin(t/2)

k=1

sinc( k ),
T

(2)

where sinc(t) = t/2 . The quintic box spline M 2 is obtained by convolving the linear box spline with itself; therefore, its Fourier transform is the square of M ( ). To fully exploit the approximation power of the box splines, they need to be combined with discrete prelters. Applying the framework of Condat and Van De Ville [Condat and Van De Ville 2007], Cs bfalvi derived an optimal discrete prelter [Cs bfalvi 2010] for e e a quasi-interpolating linear box-spline reconstruction. The Fourier transform of this prelter pM is pM ( ) = 4 1 4 cos( T ). k 3 12 k=1 (3)

The quintic box spline is not an interpolating lter, but it can be combined with a discrete prelter pM2 that makes it interpolating

[Cs bfalvi and Domonkos 2008; Finkbeiner et al. 2009; Blu et al. e 1999]. The Fourier transform of this prelter [Cs bfalvi 2010] is e pM2 ( ) = 1/s( ), where
2p

1 4 2 1 s( ) = + cos( T ) + 15 [cos(x ) + cos(y ) + cos(z )], k 5 10 k=1 (4) and = [x , y , z ]T . Note that the preltering is, in fact, a deconvolution that is performed as a division by s in the frequency domain.

-2p 2p 2p -2p

3.2

BCC B-Spline Reconstruction

Figure 2: The pass band corresponding to BCC sampling is the Voronoi cell of the dual FCC lattice, which is a rhombic dodecahedron. The red dots depict the twelve rst nearest neighbors, whereas the blue dots depict the six second nearest neighbors. the frequency-domain behavior of a lter, especially if it is nonseparable or applied on a non-separable lattice. Since, in case of BCC sampling, the spectrum of the original signal is replicated in the frequency domain around the dual FCC lattice points, the shape of the pass band is a rhombic dodecahedron, which is the Voronoi cell of the FCC lattice (see Figure 2) [Entezari et al. 2004]. To analyze the oversmoothing effect, it has to be visualized how much the frequency response of the applied reconstruction lter differs from that of the ideal low-pass lter in the pass band. Consequently, the postaliasing effect is analyzed by visualizing the frequency response in the stop band. Here the energy of the lter should be as low as possible to well approximate the ideal low-pass lter. Furthermore, it is worthwhile to emphasize those frequency components that are closer to the centers of the aliasing spectra, since mostly these components contribute to the postaliasing effect. For this purpose, we will use an appropriate penalty function. As demonstrated in Figure 1, direct volume rendering is a natural tool for visualizing an analytically known frequency response. Note that the preferred directions of the spectrum can be identied in 3D more easily than in cross-sectional 2D slices [Cs bfalvi 2010]. e Nevertheless, the images in Figure 1 show the distribution of the entire spectrum; therefore, the behavior of the lter in the pass band and in the stop band cannot be separated. First, we tried to render the spectrum using different colors inside and outside the pass band. This approach, however, was not informative when the lter had high energy in the stop band, since in this case the pass band often became almost completely hidden. Therefore, we decided to separately render the spectrum in the pass band and in the stop band. This is favorable regarding also the transfer function specication, which requires different parameter tuning for analyzing the oversmoothing and postaliasing effects.

Although the tensor-product B-splines are not tailored directly to the geometry of the BCC lattice like the linear and quintic box splines, their application on BCC-sampled data can still be reasonable because of two reasons. First, the B-splines show very good antialiasing properties on the BCC lattice [Cs bfalvi 2010]. Sece ond, a BCC B-spline ltering can be efciently implemented on the recent GPUs exploiting that the BCC lattice consists of two interleaved CC lattices [Cs bfalvi and Hadwiger 2006]. For example, e in a GPU-accelerated isosurface-rendering application, a BCC Bspline reconstruction [Cs bfalvi 2010] turned out to be four to ve e times faster than an optimized non-separable box-spline reconstruction [Finkbeiner et al. 2009] of the same approximation order. The Fourier transform of the trilinear B-spline 1 is 1 ( ) = [sinc(x )sinc(y )sinc(z )]2 . (5)

Since the tricubic B-spline 3 is the convolution of the trilinear B spline with itself, its Fourier transform is 3 ( ) = 1 ( ) 1 ( ). To better t the frequency response of the trilinear B-spline lter to the geometry of the dual FCC lattice, an optimal discrete prelter of Finite Impulse Response (FIR) has been proposed [Cs bfalvi 2010], e which guarantees a quasi-interpolation of order two. The Fourier transform of this prelter is p 1 ( ) = 5 1 4 cos( T ). k 3 6 k=1 (6)

Similarly to the trilinear B-spline, the tricubic B-spline cannot be used for interpolation on the BCC lattice either. Instead, it is suitable for quasi-interpolation of order four if it is combined with a discrete prelter p 3 of Innite Impulse Response (IIR) [Cs bfalvi e 2010]. Interestingly, the Fourier transform of this prelter is the square of pM2 . Thus, for a quasi-interpolating tricubic BCC B spline reconstruction, the same prelter is used as for the interpolating quintic box-spline reconstruction, but the preltering is performed twice.

3D Frequency-Domain Analysis

4.1

Postaliasing

Having the analytical frequency response of a reconstruction lter known, its oversmoothing and postaliasing effects can be analyzed. Marschner and Lobb dened error metrics [Marschner and Lobb 1994] for oversmoothing and postaliasing as the squared deviation from the ideal pass-band and stop-band behavior, respectively. However, in 3D, these error metrics express only the absolute measures of oversmoothing and postaliasing effects, and do not tell anything about their direction dependence or 3D distribution. Therefore, a 3D analysis is necessary to better understand

In order to visualize the stop-band behavior, we used a two-step opacity transfer function t(r( )) (see Figure 5). Function r, which 1 maps a frequency response ( ) to r( ( )) = | ( )| 6 , is responsible for making even the low-intensity parts of the spectrum visible. Function t, which enhances a characteristic isosurface of the frequency response, is dened as follows: t(r) =
1 1 8 2 [1 (1 2r) ] 1 [1 + (2r 1) 1 ] 8 2

if 0 r if
1 2

1 2

< r 1.

(7)

Linear box spline.

Trilinear B-spline.

Quintic box spline.

Tricubic B-spline.

M ( )

1 ( )

M 2 ( )

3 ( )

Frequency responses without preltering.

M ( ) pM ( )

1 ( ) p 1 ( )

M 2 ( ) pM2 ( )

3 ( ) p 3 ( )

Resultant frequency responses after a discrete preltering. Figure 3: Frequency responses of different reconstruction schemes for the BCC lattice in the stop band. The images show the interval [4 , 4 ]3 in the frequency domain. The opacity is set to zero inside the pass band, which is a rhombic dodecahedron.
transfer function for the pass band transfer function for the stop band

0.8

0.6

0.4

0.2

0.2

0.4

0.6

0.8

Figure 5: Opacity transfer functions for visualizing the frequency responses in the pass band and in the stop band.

Figure 3 shows the frequency responses only in the stop band as the opacity is set to zero inside the pass band, which is a rhombic dodecahedron. Note that the images are ordered according to the overall postaliasing effects of the corresponding reconstruction lters. The linear box spline results in the highest total postaliasing, whereas the tricubic B-spline leads to the lowest total postaliasing. Concerning the direction dependence, the box splines introduce higher aliasing frequencies along the diagonal directions, while the postaliasing effect of the B-splines is the strongest along the major axes. Note that, generally, a discrete preltering increases the total postaliasing effect but does not change the preferred directions that are mostly determined by the reconstruction lter itself.

The Fourier transform of a BCC-sampled signal contains the replicas of the primary spectrum around the points of the dual FCC lattice. Therefore, the FCC lattice points except the origin represent the DC components of the aliasing spectra, which mostly contribute to the postaliasing effect. If the frequency response of the lter is signicantly non-zero at these lattice points, a samplefrequency ripple [Marschner and Lobb 1994] might occur. Although both the box splines and the BCC B-splines ensure zerocrossings at the centers of the aliasing spectra [Entezari et al. 2008; Cs bfalvi 2010], they do not suppress the aliasing frequene cies in the vicinity of the centers uniformly along different directions. This can lead to direction-dependent postaliasing called near-sample-frequency ripple [Marschner and Lobb 1994]. Theoretically, the order of the zero-crossings (in other words, the number of vanishing moments [Entezari et al. 2008]) expresses how much the aliasing spectra are suppressed around the dual FCC lattice points. At the nearest twelve FCC lattice points (see the red dots in Figure 2), which are located along the diagonal directions at [2 , 2 , 0], [2 , 0, 2 ], and [0, 2 , 2 ], the trilinear and tricubic B-splines ensure four and eight vanishing moments, respectively [Cs bfalvi 2010]. In contrast, at these points of the e frequency domain, the linear and quintic box splines guarantee just the minimal number of their vanishing moments, which are two and four, respectively. As a consequence, based on the theory developed in [Entezari and M ller 2006], the BCC B-splines can supo press the postaliasing effect more efciently along the diagonal directions than the non-separable box splines. On the other hand, at the six second nearest FCC lattice points (see the blue dots in Figure 2) located along the major axes at [4 , 0, 0], [0, 4 , 0], and [0, 0, 4 ], the linear and quintic box splines ensure four and eight vanishing moments, while the trilinear and tricubic B-splines guarantee just the minimal number of their vanishing moments, which

Linear box spline.

Trilinear B-spline.

Quintic box spline.

Tricubic B-spline.

M ( ) w( )

1 ( ) w( )

M 2 ( ) w( )

3 ( ) w( )

Frequency responses without preltering.

M ( ) pM ( ) w( )

1 ( ) p 1 ( ) w( )

M 2 ( ) pM2 ( ) w( )

3 ( ) p 3 ( ) w( )

Resultant frequency responses after a discrete preltering. Figure 4: Frequency responses of different reconstruction schemes for the BCC lattice in the stop band. The penalty function w emphasizes those frequencies that are closer to the FCC lattice points. These points represent the DC components of the aliasing spectra, which mostly contribute to the postaliasing effect. The images show the interval [4 , 4 ]3 in the frequency domain. The opacity is set to zero inside the pass band, which is a rhombic dodecahedron.

are two and four, respectively. Thus, the non-separable box splines suppress the postaliasing effect more efciently along the major axes than the BCC B-splines. However, as the frequency responses decay rapidly, it is favorable if the number of vanishing moments is increasing by getting closer to the origin and not vice versa. The B-splines fulll this requirement better than the box splines. To emphasize the aliasing frequencies especially in the vicinity of the dual FCC lattice points, we used the following penalty function:
2

4.2

Oversmoothing

w( ) =

1 1 4 + cos( T ) k 2 8 k=1

(8)

Note that w is periodic on the FCC lattice and takes its maximum at the lattice points. Getting far from the lattice points, w is monotonically decreasing. Figure 4 shows the frequency responses in the stop band weighted by w. The images completely conrm the explanation of the direction-dependent postaliasing based on the number of vanishing moments. It is clearly apparent that the nonseparable box splines introduce aliasing frequencies in the directions of the twelve nearest FCC lattice points, while the B-splines cause aliasing in the directions of the six second nearest FCC lattice points. Considering the average magnitude of the postaliasing frequencies, the box splines perform worse than the B-splines of the same approximation powers with and without preltering. Especially around the pass band, the box splines produce a thicker postaliasing shell than the B-splines.

In order to visualize the pass-band behavior, we use only the transfer function t (see Equation 7 and Figure 5). Thus, the opacity at a frequency is set to t(| ( )|), where is the frequency response. Furthermore, we set the opacity to zero outside the rhombic dodecahedron that represents the pass band. Figure 6 shows the frequency responses rendered by these settings. Note that the lowest oversmoothing is ensured by the interpolating preltered quintic boxspline reconstruction as its frequency response almost completely lls the volume of the pass band, whereas the tricubic BCC B-spline without preltering leads to the highest oversmoothing, since most of its energy is concentrated inside a sphere of a relatively small radius. On the other hand, the IIR preltering signicantly improves the pass-band behavior of the tricubic BCC B-spline lter. Nevertheless, as the B-splines are not tailored directly to the geometry of the BCC lattice, even with preltering, they approximate the shape of the pass band with a cubic shape. In contrast, the frequency response of the interpolating preltered quintic box-spline reconstruction nicely takes the rhombic dodecahedral shape of the pass band. Generally, the discrete preltering improves the passband behavior of all the analyzed lters, but at the cost of a higher postaliasing (see Figures 3 and 4). Since the shapes of the frequency responses can be better perceived by an animated viewing, we also generated a video sequence, which is available for download at http://www.iit.bme.hu/~cseb/Publications/SCCG10/demo.wmv.

Linear box spline.

Trilinear B-spline.

Quintic box spline.

Tricubic B-spline.

M ( )

1 ( )

M 2 ( )

3 ( )

Frequency responses without preltering.

M ( ) pM ( )

1 ( ) p 1 ( )

M 2 ( ) pM2 ( )

3 ( ) p 3 ( )

Resultant frequency responses after a discrete preltering. Figure 6: Frequency responses of different reconstruction schemes for the BCC lattice in the pass band. The images show the interval [2 , 2 ]3 in the frequency domain. The opacity is set to zero outside the pass band, which is a rhombic dodecahedron.

Experimental Results

Linear box spline.

Trilinear B-spline.

To validate the results of our frequency-domain analysis, we rendered the isosurfaces of synthetic and real-world data sets using the analyzed ltering schemes. Since the linear lters cause stronger postaliasing effects than the higher-order lters, it is reasonable to apply them on a relatively oversampled volume representation. Therefore, we sampled the Marschner-Lobb signal [Marschner and Lobb 1994] on 64 64 64 2 BCC lattice points. Figure 7 shows the results of quasi-interpolating preltered linear box-spline and trilinear B-spline reconstructions. Note that the linear box spline produces artifacts exactly along the diagonal directions as it has been predicted in our 3D frequency-domain analysis. Similarly, the trilinear BCC B-spline introduces aliasing mainly along the major axes, which also conrms our theoretical hypothesis. However, on the dual FCC lattice, the second nearest aliasing spectra along the major axes are not so closed to the primary spectrum as the rst nearest aliasing spectra along the diagonal directions; therefore, the postaliasing effect of the trilinear B-spline is less apparent than that of the linear box spline. As the higher-order lters can better approximate the ideal low-pass lter than the linear lters, we tested them on a volume representation that samples the Marschner-Lobb signal closer to the Nyquist limit. Therefore, we took 32 32 32 2 BCC samples from the test signal. Figure 8 shows the results of an interpolating preltered quintic box-spline reconstruction and a quasi-interpolating preltered tricubic BCC B-spline reconstruction. The IIR preltering combined with the quintic box-spline enhances the diagonal postaliasing effects even stronger than the FIR preltering combined with the linear box spline (see Figure 7). In contrast, using

Figure 7: Quasi-interpolating preltered reconstruction of the Marschner-Lobb signal from 64 64 64 2 BCC samples using the linear box spline and the trilinear B-spline. The lower two images show the angular errors of the estimated gradients. Angular error of 30 degrees is mapped to white, while angular error of zero degree is mapped to black.

the same IIR prelter twice for a quasi-interpolating tricubic BCC B-spline reconstruction, the circular structure of the rings can be reproduced almost perfectly without introducing direction-dependent

artifacts. Although the tricubic BCC B-spline is expected to produce the highest postaliasing along the major axes, the magnitude of its aliasing effect in these directions seems to be much lower than that of the quintic box spline along the diagonal directions. Quintic box spline. Tricubic B-spline.

behavior, since its frequency response almost completely lls the rhombic dodecahedral volume of the pass band. This might explain that the quintic box spline does not blur the gradients as much as the tricubic BCC B-spline. As a consequence, the interpolating preltered quintic box-spline reconstruction leads to lower angular gradient error. Overall, combined with the known prelters, neither the BCC B-splines nor the box splines are superior, as their advantageous properties are complementary. It depends on the specic application whether the preservation of the high-frequency details or an efcient suppression of the postaliasing artifacts is considered to be more important. The non-separable box splines and the BCC B-splines are apples and oranges in this sense, since they represent different compromises between these contradictory requirements.

Acknowledgements
This work was supported by Mediso Medical Imaging Systems, the Hungarian National Ofce for Research and Technology (TECH 08/A2), the J nos a Bolyai Research Scholarship of the Hungarian Academy of Sciences, and OTKA (F68945). The rst author of this paper is a grantee of the J nos a Bolyai Scholarship.

References
Figure 8: Preltered reconstruction of the Marschner-Lobb signal from 32 32 32 2 BCC samples using an interpolating quintic box-spline ltering and a quasi-interpolating tricubic B-spline ltering. The lower two images show the angular errors of the estimated gradients. Angular error of 30 degrees is mapped to white, while angular error of zero degree is mapped to black. We compared the oversmoothing effects of the box splines and the BCC B-splines by rendering a measured data set, which is a CT scan of a carp. This data set contains thin ribs; therefore, it is especially suitable for testing the preservation of high-frequency details. In all the cases, we used comparable prelters to improve the passband behavior. Note that the box splines preserve the ribs better than the equivalent BCC B-splines, but they introduce more postaliasing artifacts that become apparent by zooming in. This behavior can also be explained by our frequency-domain analysis, since the box splines can approximate the ideal low-pass lter better in the pass band but worse in the stop band.
B ENTUM , M. J., L ICHTENBELT, B. B. A., AND M ALZBENDER , T. 1996. Frequency analysis of gradient estimators in volume rendering. IEEE Transactions on Visualization and Computer Graphics 2, 3, 242254. B LU , T., T H E VENAZ , P., AND U NSER , M. 1999. Generalized interpolation: Higher quality at no additional cost. In Proceedings of IEEE International Conference on Image Processing, 667671. C ONDAT, L., AND VAN D E V ILLE , D. 2007. Quasi-interpolating spline models for hexagonally-sampled data. IEEE Transactions on Image Processing 16, 5, 11951206. C S E BFALVI , B., AND D OMONKOS , B. 2008. Pass-band optimal reconstruction on the body-centered cubic lattice. In Proceedings of Vision, Modeling, and Visualization, 7180. C S E BFALVI , B., AND H ADWIGER , M. 2006. Preltered B-spline reconstruction for hardware-accelerated rendering of optimally sampled volumetric data. In Proceedings of Vision, Modeling, and Visualization, 325332. C S E BFALVI , B. 2005. Preltered Gaussian reconstruction for high-quality rendering of volumetric data sampled on a body-centered cubic grid. In Proceedings of IEEE Visualization, 311318. C S E BFALVI , B. 2008. BCC-splines: Generalization of B-splines for the body-centered cubic lattice. Journal of WSCG 16, 13, 8188. C S E BFALVI , B. 2008. An evaluation of preltered reconstruction schemes for volume rendering. IEEE Transactions on Visualization and Computer Graphics 14, 2, 289301. C S E BFALVI , B. 2010. An evaluation of preltered B-spline reconstruction for quasi-interpolation on the body-centered cubic lattice. IEEE Transactions on Visualization and Computer Graphics 16, 3, 499512. E NTEZARI , A., AND M OLLER , T. 2006. Extensions of the Zwart-Powell box spline for volumetric data reconstruction on the Cartesian lattice. IEEE Transactions on Visualization and Computer Graphics (Proceedings of IEEE Visualization) 12, 5, 13371344. E NTEZARI , A., DYER , R., AND M OLLER , T. 2004. Linear and cubic box splines for the body centered cubic lattice. In Proceedings of IEEE Visualization, 1118. E NTEZARI , A., VAN D E V ILLE , D., AND M OLLER , T. 2008. Practical box splines for reconstruction on the body centered cubic lattice. IEEE Transactions on Visualization and Computer Graphics 14, 2, 313328.

Conclusion

In this paper, we have presented the rst 3D frequency-domain analysis of reconstruction lters previously proposed for the optimal BCC lattice. According to our results, both the box splines and the BCC B-splines cause direction-dependent postaliasing effects, but along different directions. The box splines maximize the postaliasing along the diagonal directions, while the BCC B-splines produce the highest postaliasing along the major axes. However, the average magnitude of the aliasing frequencies is lower in the frequency responses of the BCC B-splines with and without preltering. The B-spline lters suppress especially the nearest aliasing spectra more efciently than the box-spline lters. Therefore, the B-splines were able to reconstruct the circular structure of the rings in the Marschner-Lobb signal much more isotropically with minimal direction-dependent artifacts. In contrast, the box splines perform better in the pass band; therefore, they can capture more highfrequency details. Particularly, the interpolating preltered quintic box-spline reconstruction can well approximate the ideal pass-band

Quasi-interpolating preltered linear box-spline reconstruction.

Quasi-interpolating preltered trilinear B-spline reconstruction.

Interpolating preltered quintic box-spline reconstruction.

Quasi-interpolating preltered tricubic B-spline reconstruction. Figure 9: Reconstruction of the skeleton of a carp from http://www9.informatik.uni-erlangen.de/External/vollib). 128 128 256 2 BCC samples (courtesy of

E NTEZARI , A., M IRZARGAR , M., AND K ALANTARI , L. 2009. Quasiinterpolation on the body centered cubic lattice. Computer Graphics Forum (Proceedings of EuroVis) 28, 3, 10151022. F INKBEINER , B., E NTEZARI , A., M OLLER , T., AND VAN D E V ILLE , D. 2009. Efcient volume rendering on the body centered cubic lattice using box splines. TR-CMPT2009-04, Graphics, Usability and Visualization (GrUVi) laboratory, Simon Fraser University. M ARSCHNER , S., AND L OBB , R. 1994. An evaluation of reconstruction lters for volume rendering. In Proceedings of IEEE Visualization, 100 107. M ATTAUSCH , O. 2003. Practical reconstruction schemes and hardwareaccelerated direct volume rendering on body-centered cubic grids. Masters Thesis, Vienna University of Technology. M ITCHELL , D., AND N ETRAVALI , A. 1988. Reconstruction lters in computer graphics. In Proceedings of SIGGRAPH, 221228. M OLLER , T., M UELLER , K., K URZION , Y., M ACHIRAJU , R., AND YAGEL , R. 1998. Design of accurate and smooth lters for function

and derivative reconstruction. In Proceedings of IEEE Symposium on Volume Visualization, 143151. O PPENHEIM , A. V., AND S CHAFER , R. W. 1989. Discrete-Time Signal Processing. Prentice Hall Inc., Englewood Cliffs, 2nd edition. S IGG , C., AND H ADWIGER , M. 2005. Fast third-order texture ltering. In GPU Gems 2: Programming Techniques for High-Performance Graphics and General-Purpose Computation, Matt Pharr (ed.), AddisonWesley, 313329. T HEUSSL , T., H AUSER , H., AND G R OLLER , M. E. 2000. Mastering windows: Improving reconstruction. In Proceedings of IEEE Symposium on Volume Visualization, 101108. T HEUSSL , T., M OLLER , T., AND G R OLLER , M. E. 2001. Optimal regular volume sampling. In Proceedings of IEEE Visualization, 9198. VAN D E V ILLE , D., B LU , T., AND U NSER , M. 2004. Hex-splines: A novel spline family for hexagonal lattices. IEEE Transactions on Image Processing 13, 6, 758772.