Interactively Controlling the Smoothing and Postaliasing Eects in Volume Visualization

Bal zs Cs bfalvi a e Bal zs Domonkos a Department of Control Engineering and Information Technology Budapest University of Technology and Economics

(a)

(b)

(c)

Figure 1: Reconstruction of the Marschner-Lobb signal using our scalable volume-rendering technique. (a) Maximal smoothing and minimal postaliasing. (b) Trade-off between smoothing and postaliasing. (c) Minimal smoothing and maximal postaliasing.

Abstract

Introduction

In volume-rendering applications an appropriate resampling lter is usually chosen by making a compromise between quality and efciency. Generally, the ne details can be reconstructed by lters of wider support that can better approximate the ideal low-pass lter. On the other hand, if the data is noisy, a lter of a good pass-band behavior might even emphasize the noise. Therefore, to visualize noisy data, a lter of a higher smoothing effect is more favorable. Thus, the choice of the reconstruction lter depends on the quality of the data and the purpose of the visualization as well. In this paper, we propose a scalable volume-rendering technique for interactively controlling the frequency-domain behavior of the reconstruction. Applying our method, the trade-off between the smoothing and postaliasing effects can be set on the y by using a single slider.

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

Following a signal-processing approach, direct volume rendering can be interpreted as a resampling of a 3D discrete function. If the values of this discrete function are regular samples of an originally continuous signal, the resampling can be easily performed by a convolution with a continuous reconstruction kernel. Theoretically, a band-limited continuous signal sampled above the Nyquist limit can be perfectly reconstructed by convolving the discrete samples with the ideal sinc lter. The sinc lter, however, is impractical because of its innite support. Therefore, it is usually approximated by lters of Finite Impulse Response (FIR) in practice [Theul et al. 2000]. Due to the deviation from the ideal pass-band and stop-band behavior, practical lters introduce oversmoothing and postaliasing [Marschner and Lobb 1994]. Generally, the wider the support of the lter, the better it can approximate the ideal frequency-domain behavior. On the other hand, a lter of a wider support is more expensive to evaluate. Moreover, a better approximation of the ideal low-pass lter does not necessarily improve the reconstruction quality if the data is noisy or undersampled. In this case a lter of a stronger smoothing effect might be preferred. As the choice of an appropriate lter depends on the data set and on the specic visualization task as well, a universal solution does not exist. Therefore, in this paper, we propose an interactive volume-rendering technique that offers a continuous transition between different reconstructions controlled by a single parameter (see Figure 1). Since the user gets an immediate visual feedback whenever this parameter is modied,

Keywords: volume rendering, preltered reconstruction, generalized interpolation

e-mail: e-mail:

cseb@iit.bme.hu domonkos@ik.bme.hu

he/she does not need a priori knowledge about the smoothing and aliasing effects of the different ltering schemes. A brute-force implementation of such a scalable volume rendering would use two reconstruction lters with signicantly different frequency-domain behavior and a weighted sum of the results would be calculated. In this case the bottleneck would be the more expensive lter. In contrast, we use the same reconstruction lter twice but on different data values. The rst data value is the original one and the second data value is obtained by a discrete preltering. Exploiting the SIMD instructions of the GPU, these two data values can be convolved with the reconstruction kernel at the same time, and a weighted sum of the results can be efciently calculated. Currently the trilinear reconstruction lter is considered to be a de facto standard for volume rendering as it is supported by the recent GPUs. Although the tricubic lters are more expensive computationally, they provide signicantly higher image quality due to their higher continuity and approximation orders [M ller et al. 1998]. o Recently, it has been shown that a tricubic B-spline ltering can also be performed interactively utilizing the fast trilinear texturefetching capability of the GPU [Sigg and Hadwiger 2005]. Additionally, it is well known that the approximation power of the tricubic B-spline can be fully utilized only if the discrete samples are preltered before the convolution with the continuous reconstruction kernel [Condat et al. 2005]. Such a preltering is not unique. Different prelters lead to different resultant frequency responses [Cs bfalvi 2008]. Although optimal prelters have been proposed e for the trilinear kernel as well [Blu et al. 2004; Condat et al. 2005], according to our experience, they do not change the frequencydomain behavior as much as for the tricubic B-spline ltering. Therefore, in our implementation we have chosen the tricubic Bspline as a reconstruction kernel and a preltering for generalized interpolation [Blu et al. 1999]. This preltering scheme ensures a quite good pass-band behavior for a slightly increased postaliasing effect. In contrast, the tricubic B-spline without preltering results in signicant oversmoothing and relatively low postaliasing. In Section 2, the previous work on reconstruction for volume rendering is reviewed. We briey summarize the preltered reconstruction scheme in Section 3. In Sections 4 and 5, we discuss the theoretical differences between the approximating cubic B-spline and the interpolating preltered cubic B-spline reconstructions. In Section 6, a continuous transition between these two reconstructions is proposed, which is interactively controlled by a free parameter. In Section 7, we present examples of practical volume visualization, where such a continuous transition makes sense, and helps the user distinguish the noise from the useful content of the data. The GPU implementation is reported in Section 8, and nally in Section 9 the contribution of this paper is summarized.

discrete prefilter

continuous postfilter

fk

pk

ck w(t)

(t)

~ f(t)

prefiltered reconstruction

Figure 2: Preltered reconstruction: Input samples fk are rst convolved with a discrete prelter pk yielding coefcients ck , which are then convolved with a continuous postlter (t). The resultant impulse response is denoted by w(t).

A reconstruction lter can also be analyzed based on the approximation theory. Here the major aspect is how fast the approximate signal converges to the original one by decreasing the sampling distance. This mainly depends on the approximation power of the reconstruction lter. The order of approximation is L if the frequency response equals to zero at the centers of all the aliasing spectra with a multiplicity of at least L [Strang and Fix 1971]. However, in order to fully exploit the approximation power of a given lter, usually an appropriate discrete preltering is necessary (see Figure 2). Such a preltering can ensure that a polynomial of L 1 or lower degree is perfectly reconstructed. If this condition is satised, the reconstruction scheme is quasi-interpolating of order L [Condat et al. 2005].

Using the lter design approach of M ller et al. [M ller et al. 1998], o o piecewise polynomial lters that are quasi-interpolating of a certain order can be directly derived. The coefcients of the polynomials are determined by solving a linear equation system such that the required accuracy and smoothness criteria are satised. A lter is classied as k-EF if the Error Function of the approximation is of order k [M ller et al. 1996], that is, the reconstruction is quasio interpolating of order k. For example, if a piecewise polynomial lter of minimal support is searched for, which guarantees a C2 continuous 2-EF reconstruction then the classical cubic B-spline is obtained. However, without preltering the approximation power of the cubic B-spline, that is L = 4, is not fully exploited. Therefore Cs bfalvi and Domonkos [Cs bfalvi and Domonkos 2009] exe e tended the lter design by taking also a discrete preltering into account. Using their method, the continuous reconstruction lter is xed and a discrete prelter of minimal support is searched for which maximizes the order of accuracy.

Related Work

For high-quality volume rendering, an appropriate choice of the reconstruction lter is of crucial importance. Therefore many researchers analyzed practical reconstruction lters from different aspects. From a signal processing point of view, the accuracy of the reconstruction is mainly inuenced by the global frequency-domain behavior of the applied lter, especially if the original signal is sampled near the Nyquist limit. A lter is usually characterized by a frequency plot and its quality is quantitatively evaluated as the deviation from the ideal pass-band and stop-band behavior [Marschner and Lobb 1994; Bentum et al. 1996]. The main drawback of this approach is that practical signals can hardly be considered bandlimited. Thus even the sinc lter produces ringing artifacts due to the drastic cut-off in the frequency domain [Artner et al. 2005].

Generally, a discrete preltering can improve the reconstruction from other aspects as well. For example, depending on the applied discrete prelter, it can optimize the pass-band behavior of the reconstruction [Malzbender 1993], make a non-interpolating continuous postlter interpolating [Blu et al. 1999; Th venaz et al. 2000; e Blu et al. 2004], or increase the accuracy of the reconstruction in a sense of minimal approximation error [Condat et al. 2005]. All these preltering techniques are of Innite Impulse Response (IIR) and proven to yield k-EF reconstruction if the approximation order of the continuous postlter is L = k [Cs bfalvi 2008]. This implies e that the result is a quasi-interpolation of order k. Practically, the order of accuracy becomes important if the sampling frequency is increased, so the original signal gets oversampled. This assumption is acceptable for medical volume rendering as the resolution of CT and MRI scanners has been signicantly increased in the last two decades.

cubic B-spline 1 1

cubic B-spline interpolating prefiltered cubic B-spline ideal low-pass filter

0.8

0.8

0.6

0.6

0.4

0.4

0.2

0.2

-2

-1.5

-1

-0.5

0.5

1.5

0.5

1.5

Figure 3: Impulse response of the approximating cubic B-spline lter.

Figure 4: Frequency response of the cubic B-spline with and without preltering for generalized interpolation (the horizontal axis represents the ordinary frequency = 2 ).

Preltered Reconstruction
cubic B-spline provides just a 2-EF function reconstruction [M ller o et al. 1998]. Moreover, the cubic B-spline lter is not interpolating, just approximating as 3 (1) = 0. Due to its relatively strong oversmoothing effect, it is mainly used for reconstructing data which is corrupted by noise. Theoretically the tricubic B-spline lter kernel covers 64 voxels, so the calculation of a sample at an arbitrary position would require 64 nearest-neighbor texture fetches. However, using the acceleration technique proposed by Sigg and Hadwiger [Sigg and Hadwiger 2005], a tricubic B-spline ltering can be evaluated also from eight trilinear texture samples. Since the recent GPUs can perform nearest-neighbor and trilinear texture fetches in approximately the same time, this method is about eight times faster than the bruteforce implementation, and can be used for interactive high-quality volume rendering.

According to the traditional approach, a continuous function f (t) can be reconstructed from its evenly located samples fk = f (kT ) by the following convolution: f (t) f(t) =

k=

fk (

t k), T

(1)

where f(t) is the approximate reconstruction of f (t), T is the sampling distance, and (t) is a continuous reconstruction lter. In contrast, using preltered reconstruction (see Figure 2), the lter kernel is convolved with coefcients ck , which are not necessarily equal to the samples fk : f(t) =

k=

ck (

t k). T

(2)

Coefcients ck are determined by convolving the original samples fk with a discrete prelter pk : ck =

j=

Interpolating Preltered Cubic B-Spline Reconstruction

fk j p j .

(3) Although the cubic B-spline lter is not interpolating, applying the principle of generalized interpolation [Blu et al. 1999], it can still be used for interpolation. In this case, the coefcients ck of the shifted cubic B-splines are determined by a discrete preltering of the original samples fk (see Figure 2). The preltering is basically a discrete deconvolution with the sampled cubic B-spline kernel: ck = fk ( 3 (k))1 . Due to the deconvolution the resultant impulse response is of IIR, so the ideal lowpass lter can be better approximated. The deconvolution can be easily implemented as a division in the frequency domain. The resultant frequency response of the interpolating preltered cubic B spline reconstruction is 3 ( )/DTFT[ 3 (k)]( ) = 3 ( ) 3/(2 + e cos( )) [Cs bfalvi 2008]. Figure 4 shows this frequency response compared to that of the non-preltered cubic B-spline reconstruction. Note that the preltering signicantly improves the pass-band behavior at a cost of slightly higher postaliasing. The preltering improves the numerical accuracy as well, since it e results in a C2 4-EF reconstruction [Cs bfalvi 2008] optimally utilizing the approximation power of the cubic B-spline. Thus, generalized interpolation applied for the cubic B-spline satises also the condition for quasi-interpolation of order 4.

Approximating Cubic B-Spline Reconstruction

The cubic B-spline lter is dened as follows (see Figure 3): 1 |t|3 |t|2 + 2 if |t| 1 2 3 3 (t) = 1 |t|3 + |t|2 2|t| + 4 if 1 < |t| 2 6 3 0 otherwise.

(4)

The Fourier transform of the cubic B-spline is 3 ( ) = sinc4 ( ), sin( /2) where sinc( ) = /2 [Th venaz et al. 2000]. Note that 3 e guarantees zero-crossings of order 4 at the centers of all the aliasing spectra (if = j2 , where j Z \ {0}), therefore the approximation power of the cubic B-spline is L = 4. As a consequence, there exists a discrete prelter for the cubic B-spline that ensures a C2 continuous 4-EF function reconstruction, or in other words, a quasi-interpolation of order 4. Without preltering, however, the

Reconstruction from 40 40 40 voxels.

Reconstruction from 60 60 60 voxels.

Tricubic B-spline.

Interpolating tricubic B-spline.

Tricubic B-spline.

Interpolating tricubic B-spline.

Figure 5: Reconstructions of the Marschner-Lobb signal from 40 40 40 (rst two columns) and from 60 60 60 (second two columns) voxels. First row: shaded isosurface rendering. Second row: reconstruction of a 2D slice. Third row: angular error between the original gradient and the gradient of the reconstructed signal. Angular error of zero degree is mapped to black, whereas angular error of 30 degrees is mapped to white.

Transition Between Approximation and Interpolation

The characteristic behavior of a reconstruction scheme can be well demonstrated on the classical Marschner-Lobb signal [Marschner and Lobb 1994; Bentum et al. 1996]. However, concerning the ranking of the different schemes, the interpretation of the images is usually rather subjective in the literature. Some researchers prefer the preservation of high-frequency details, while others prefer a visually pleasing reconstruction with smooth surfaces and minimal aliasing. In this paper, we would like to let the potential user decide which scheme to prefer. To help this decision, our method allows a continuous transition between two rather different reconstructions, which are the preltered and non-preltered cubic B-spline reconstructions. Figure 5 shows the differences on the Marschner-Lobb signal. The cubic B-spline without preltering leads to severe oversmoothing but it does not introduce so much postaliasing. The oversmoothing effect is apparent especially on the cross-sectional slice. In contrast, the preltered cubic B-spline reconstruction performs much better in reconstructing the high-frequency details like the depth and height of the outer waves. Nevertheless, the price for that is a higher postaliasing. Generally, forcing a higher-degree approximate function to pass exactly through the data values causes oscillation or overshooting [Marschner and Lobb 1994] if the data is not sampled at a sufcient sampling frequency. On the other hand, the preltering ensures clearly higher numerical accuracy if the data values are not corrupted by measurement and quantization noise and the data is assumed to be sampled at least near the Nyquist limit. In practice this is usually not the case.

Our method evaluates the two different reconstructions and calculates their weighted average controlled by a free parameter r. If r = 0, the preltered values are not taken into account, which results in the standard C2 2-EF approximating tricubic B-spline reconstruction. If r = 1 then only the preltered values are considered, which leads to the C2 4-EF interpolating preltered tricubic B-spline reconstruction. If the value of r is in between, a linear transition between the two reconstructions is calculated. Due to the efcient GPU implementation, parameter r can be controlled interactively ensuring an immediate visual feedback. We think that using such an interactive control of the trade-off between approximation and interpolation, smoothing and postaliasing, or noise-suppression and detail-enhancement, the characteristic effect of the given reconstruction scheme can be immediately understood even by a user who is not necessarily an expert in sampling theory. For example, assume that a physician wants to detect a stenosis in a CT scan. An inappropriate resampling can easily lead to a false diagnosis if the applied lter has a stronger smoothing effect and therefore blood vessels get disconnected. On the other hand, point-like noise is usual in CT scans, which can be suppressed by a smoothing lter. This is a typical case, when none of the extreme values of r leads to a satisfactory result and it is worthwhile to play around with parameter r to clearly separate the information that is contained in the data from the effect of the reconstruction scheme. The continuous transition rendered in one window is denitely more useful than separately rendering two images by the two reconstruction schemes and searching for the differences. Our method automatically attracts the eyes of the user to those regions, where the two reconstruction techniques make a signicant difference. For example, a blood vessel, which is disconnected due to the stronger smoothing effect of the pure tricubic B-spline reconstruction, gets

progressively connected if the pass-band behavior is improved by increasing parameter r. If the rendered image does not change drastically when parameter r is modied than the user can be sure that the data has been sufciently sampled and the result does not depend that much on the applied resampling scheme.

Reconstruction scheme
Classical cubic B-spline ltering Our scalable cubic B-spline ltering

CT carp 3.01 fps 2.74 fps

CT head 6.22 fps 5.65 fps

MRI angio 7.42 fps 6.78 fps

Table 1: The frame rates of texture-based isosurface rendering using the classical cubic B-spline ltering and our scalable cubic Bspline ltering.

Experimental Analysis
the trilinear samples, the continuous transition between them is obtained as a result of a lerp instruction parametrized by the current value of r. The modied scalable cubic B-spline ltering is just slightly slower than the traditional one. Table 1 shows the rendering rates for our test data sets. The isosurfaces are rendered applying a simple texture-slicing approach [Westermann and Ertl 1998; Cabral et al. 1994] without optimization like empty space leaping. In average, the price for the scalability is a loss of about 10% in the rendering performance. Note that our method can be easily adapted to texture-based ray casting [Kr ger and Westermann 2003] as well. u Furthermore, it can be combined also with more sophisticated visualization models, like semitransparent volume rendering.

Using our scalable volume-rendering technique we rendered different real-world data sets. Figure 6 shows the results for different values of parameter r. The rst test data set is a CT scan of a carp, which consists of 256 256 512 voxels. The images well demonstrate the importance of the discrete preltering even if the data is sampled at a relatively high frequency. The ne ribs of the carp can be best reconstructed if parameter r equals to 1. In this case, a noise suppression by a smoother reconstruction (r < 1) is not necessary as the data does not seem to be noisy. The second data set is a CT scan of a human head, which contains 256 256 159 voxels. The blood vessels are enhanced due to the injected contrast agent. Using the pure cubic B-spline ltering (r = 0), the blood vessels get disconnected because of the stronger smoothing effect. Although the interpolating preltered cubic Bspline reconstruction (r = 1) nicely connects the blood vessels, it enhances the point-like noise as well. Here the noise is introduced mainly because of prealiasing, since the sampling frequency in the data acquisition was not sufcient to represent even the nest vessel structures. However, an interactive control of parameter r can help the user in separating the useful content from the noise. The images of the third test data set, which is an MRI angiography data of resolution 256 256 124, also justify the importance of an interactive control between smoothing and detail-enhancement. In this example, the resolution of the original data is not sufcient either, so even the preltered cubic B-spline resampling has difculties in reconstructing the capillary vessels. The inappropriate reconstruction of the ne structures visually overloads the image with too many noisy details, which is more annoying than useful. Therefore, it is worthwhile to set an appropriate value of parameter r, to clearly see at least the thicker parts of the blood vessels by removing the noise.

Conclusion and Future Work

In this paper, a scalable volume-rendering technique has been presented. Using our method, the trade-off between the smoothing and postaliasing effects can be interactively controlled. This is helpful especially if the data contains high-frequency noise as well as ne details to be reconstructed. The user does not have to know the frequency-domain behavior of the different reconstruction schemes in advance, since the interactive parameter control provides an immediate visual feedback. In our future work, we plan to use our method adaptively by setting parameter r with an appropriate transfer function, such that it can vary inside the volume depending on the density, the gradient magnitude, or the location. For example, near the region of interest the value of r can be increased to reconstruct the ne details, whereas in the unimportant regions r is reduced to remove the noise from the context.

Implementation Acknowledgements

We implemented our method in OpenGL and Cg and tested it on an NVIDIA GeForce 6800 GT graphics card. Our implementation is based on the efcient cubic B-spline ltering proposed by Sigg and Hadwiger [Sigg and Hadwiger 2005]. The main difference is that we store not just the original data value for each voxel but a preltered value for generalized interpolation as well. Therefore we use texture format GL LUMINANCE ALPHA16F ARB, which provides 16-bit oating-point precision. Assuming that in the original data each voxel is stored in 16 bits (which is usual in medical data sets), our method doubles the storage requirements. The original fast cubic B-spline ltering consists of two steps. The rst step is the calculation of the texture coordinates for the trilinear texture fetches. In the second step a weighted sum of the trilinear texture samples is calculated. The rst step is the same in our implementation, whereas in the second step we exploit the SIMD architecture of the GPU to trilinearly interpolate the original (GL LUMINACE) and the preltered (GL ALPHA) data values at the same time. After having the tricubic samples calculated from

This work was supported by the J nos Bolyai Research Scholarship a of the Hungarian Academy of Sciences, OTKA (F68945), the Hungarian National Ofce for Research and Technology, and HewlettPackard. The rst author of this paper is a grantee of the J nos a Bolyai Scholarship.

References
A RTNER , M., M OLLER , T., V IOLA , I., AND G R OLLER , M. E. 2005. High-quality volume rendering with resampling in the frequency domain. In Proceedings of Joint EUROGRAPHICSIEEE VGTC Symposium on Visualization (EuroVis), 8592. 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, 667 671. B LU , T., T H E VENAZ , P., AND U NSER , M. 2004. Linear interpolation revitalized. IEEE Transactions on Image Processing 13, 5, 710719. C ABRAL , B., C AM , N., AND F ORAN , J. 1994. Accelerated volume rendering and tomographic reconstruction using texture mapping hardware. In Proceedings of IEEE Symposium on Volume Visualization, 9198. C ONDAT, L., B LU , T., AND U NSER , M. 2005. Beyond interpolation: Optimal reconstruction by quasi-interpolation. In Proceedings of International Conference on Image Processing (ICIP), 3336. C S E BFALVI , B., AND D OMONKOS , B. 2009. Preltered gradient reconstruction for volume rendering. In Proceedings of Winter School of Computer Graphics. 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. K R UGER , J., AND W ESTERMANN , R. 2003. Acceleration Techniques for GPU-based Volume Rendering. In Proceedings IEEE Visualization, 3845. M ALZBENDER , T. 1993. Fourier volume rendering. ACM Transactions on Graphics 12, 3, 233250. M ARSCHNER , S., AND L OBB , R. 1994. An evaluation of reconstruction lters for volume rendering. In Proceedings of IEEE Visualization, 100107. M OLLER , T., M ACHIRAJU , R., M UELLER , K., AND YAGEL , R. 1996. Classication and local error estimation of interpolation and derivative lters for volume rendering. In Proceedings of IEEE Symposium on Volume Visualization, 7178. 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. S IGG , C., AND H ADWIGER , M. 2005. Fast third-order texture ltering. In GPU Gems 2: Programming Techniques for HighPerformance Graphics and General-Purpose Computation, Matt Pharr (ed.), Addison-Wesley, 313329. S TRANG , G., AND F IX , G. 1971. A Fourier analysis of the nite element variational method. In Constructive Aspects of Functional Analysis, 796830. 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 H E VENAZ , P., B LU , T., AND U NSER , M. 2000. Interpolation revisited. IEEE Transactions on Medical Imaging 19, 7, 739 758. W ESTERMANN , R., AND E RTL , T. 1998. Efciently using graphics hardware in volume rendering applications. Computer Graphics (Proceedings of SIGGRAPH 98), 169176.

Non-preltered tricubic B-spline reconstruction (r = 0).

Transition between non-preltered and preltered tricubic B-spline reconstructions (r = 0.5).

Preltered tricubic B-spline reconstruction (r = 1).

CT scan of a carp (256 256 512 voxels).

CT scan of a human head (256 256 159 voxels).

MRI angiography data (256 256 124 voxels). Figure 6: Volume rendering of different real-world data sets. First column: Non-preltered tricubic B-spline reconstruction (r = 0). Second column: Transition between non-preltered and preltered tricubic B-spline reconstructions (r = 0.5). Third column: Preltered tricubic B-spline reconstruction (r = 1). Generally, by increasing the contribution of the preltered data values the ne details can be better captured but the high-frequency noise contained in the data (see the visualization of the MRI data in the last row) might be enhanced as well. A preferred trade-off can be set by interactively controlling parameter r.