Proceedings of the IIEEJ Image Electronics and Visual Computing Workshop 2012 Kuching, Malaysia, November 21-24, 2012 A Rendering Method for Subsurface Scattering Effects Using Interpolated Luminance Distribution Functions † ‡ ‡ ‡ ‡ Tomohisa MANABE , Ayaka FURUTSUKI , Bisser RAYTCHEV , Toru TAMAKI and Kazufumi KANEDA †Kagawa National College of Technology, ‡Graduate School of Engineering, Hiroshima University ABSTRACT In this paper, we propose a method for rendering translucent objects which takes into account subsurface scattering phenomena. To calculate the luminance of a translucent object, we introduce a convolution-based approach which uses the luminance distributions caused by a set of incident light beams to the object surface. The luminance distributions of the light beams are calculated from a pair of sampled distributions. Depending on the incident angle of the light beam and the material parameters of the translucent object, the luminance distributions are interpolated in an angular parameter space obtained from the original planar space. We also propose an efficient method for calculating the convolution using light beams of different sizes. Several rendered images of translucent objects lit by slanting light beams demonstrate the usefulness of the proposed method. 1. INTRODUCTION In order to be able to render realistically translucent objects (like human skin or marble, etc.), subsurface scattering needs to be taken into consideration. Several methods able to render subsurface scattering have been previously developed. Among them, the dipole light source model [1] provides a simple method to generate realistic images and is widely used in many applications. However, the luminance distributions generated by the dipole model are different from the actual luminance distributions, especially in the case when a light beam hits a translucent object in an oblique direction. Figure 1 shows a luminance distribution taken from a real photograph. An acrylic plastic plate is lit by a laser beam at 45 degrees incident angle (Figure 1 (a)). In Figure 1 (b) pseudo-color is assigned to the luminance values to easily observe the luminance distribution around the light incidence point. The brightest point shifts from the incident point to the right that is the inclined direction of the beam. The luminance distribution is also deformed into an elliptical shape. Traditional methods cannot render such phenomena of subsurface scattering. We propose a method for rendering subsurface scattering effects taking into account the change of luminance distributions depending on both the incident angle of light and the material parameters of a translucent object. The proposed method uses a convolution approach to calculate the luminance of a translucent object. That is, light cast onto the surface of the translucent object is decomposed into thin beams, and the luminance distributions generated by the beams are accumulated to calculate the luminance on the translucent object. We also propose a method for accelerating the calculation of the convolution using an adaptive size of the decomposed beams. In the proposed method, the luminance distributions caused by the light beams are interpolated from a couple of luminance distributions depending on the incident angle of the beam. Taking into account the property of the luminance distributions caused by slanting light beams, the interpolation of the luminance distributions is executed in an angular parameter space converted from an original planar space. The method is also able to interpolate luminance distributions depending on the material parameters of a translucent material. We need only few distribution samples to generate various kinds of luminance distributions for light beams from any direction and for any material parameters. A translucent material with different albedos for different wavelengths of light can be rendered by the proposed method. Also, subsurface scattering phenomena depending on the incident angle of light are successfully rendered. The rendered images of a translucent object lit by slanting light beams demonstrate the usefulness of the proposed method. (a) Photograph (b) Pseudo-color image Figure 1 The luminance distribution of an acrylic plastic plate lit by an inclined light beam 2. RELATED WORK Several methods have been developed to render a translucent object. An applicative model using a dipole point light source for approximating the components of multiple scattering was developed in [1]. As dipole, a positive virtual source is set beneath the incident point and a negative virtual source above the point. The dipole model can display translucent objects which have a smooth and soft appearance. A faster version of this method has been developed in [2], while Mertens et al. further made possible the interactive rendering of transformable semitransparent materials [3]. Chen et al. proposed a method for relating a texture function and dipole light sources to a photon mapping [4]. Donner and Jensen [5] proposed to use a multipole light source for rendering multilayer materials (such as human skin) and thin materials (like paper). Because all these methods simply approximate the multiple scattering’s components through models based on experience, they are unable to account for some actual phenomena like, for example, when the brightest point of the luminance distribution moves away from the incident point of a beam and the luminance distribution gets distorted. To solve the problems with the dipole point source and the multipole point source models, Masuike et al. designed a simulation model based on the theory of light scattering [6]. However, when the secondary scattering of light is considered this method is computationally too expensive and cannot be used directly for display. Takamura et al. proposed a method for interpolating the intermediate irradiance distributions using the SubSurface Scattering Irradiance Distribution (SSSID) obtained from the simulation when the incident angle of light or material parameters are changed [7]. The method was further improved to be able to render complex occlusion patterns efficiently [8]. Shinya et al. proposed a method for finding a semianalytical solution by using plane-parallel approximation in a volume rendering equation [9]. Plane-parallel approximation is simplified by assuming a directional light [10]. Moreover, it was shown that plane-parallel approximation can be applied to anisotropic scattering, and was used in a model for rendering hairy objects [11]. Donner et al. proposed a rendering method using a BSSRDF based on a simulation of photon mapping [12]. In the method, an experimental BSSRDF of a simulated material is constructed. Hence, it is necessary to construct BSSRDF by preparing a database of irradiance distributions when the albedo or phase function is different. Here, we further extend Takamura’s method [7] to be able to render subsurface scattering effects such as the shift of the brightest point and the distorted luminance distributions mentioned above. The proposed method interpolates the luminance distribution for both the incident angle of light and the material parameters simultaneously. The improvement makes it possible to render various kinds of translucent materials as well as spectral (color) images. 3. RENDERING SUBSURFACE SCATTERING USING INTERPOLATED LDFS 3.1. Outline of the proposed method The proposed method is based on a convolution approach to calculate the luminance of a translucent material. Given the impulse response function g of a system, the output L to an input signal f is calculated by the following convolution equation: . (1) When calculating the luminance of a surface, we need a 2D space (u, v) and the response function changes depending on both the incident angle, θ, of the light beam and the material parameters, α, of the translucent object, such as the scattering coefficients, albedos, phase functions, etc. Considering these peculiarities of subsurface scattering, we derive the following equation for the luminance calculation (see Figure 2): , , , , , (2) where the input signal f(u, v) represents the intensity of the incident beam at point (u, v) on the surface. The response function gθ,α is the luminance distribution when a single beam with incident angle θ hits a translucent object with material parameters α, and will be called the Luminance Distribution Function (LDF) in the following discussion. We calculate the convolution in Eq. 2 over the area of the object surface to get the luminance L of the surface. Then, the luminance L is mapped onto the surface to display the projection image. response v incident angle θ beam v u u material parameter α gθ, α(u,v) (a) Luminance Distribution Function (LDF) incident beams y L(x, y) v u v u v y x u (b) Convolution to calculate the luminance Figure 2 Luminance calculation x 3.2. Interpolation of LDFs When the incident angle of a light beam is changed and/or the material parameters of a translucent object are changed, the LDF also changes. To get the LDFs corresponding to the incident light beams and the material parameters, we introduce an interpolation scheme using a couple of the LDF samples whose incident angle and material parameters are known. Note that a linear interpolation in the original planar space (u, v) does not give a good LDF, because the luminance value at a specified point P(u, v) changes nonlinearly. To address this problem, we transform the original planar space to an angular space (φ, ψ) as shown in Figure 3. First, we set a “reference point” Q above the brightest point H of the LDF sample. The azimuth angle φ is defined as the angle between vectors PO and PH, which originate from P(u, v) and are oriented in the directions of the beam incident point O and the brightest point H, respectively. The zenith angle ψ is defined as the angle between the surface normal n and the vector PQ pointing to the reference point Q. Figure 4 shows LDFs in the original planar space (u, v) and in the angular space (φ, ψ). As shown in Figure 4(a), the LDF in the original planar space has a radial luminance distribution whose center is the brightest point, and the attenuation rates of the luminance values are differ in each direction. Figure 4 (b) shows the LDF in the angular space (φ, ψ) based on the reference point. The luminance values decrease monotonically in the azimuth angle φ, and vary gently in a limited range in the zenith angle ψ. We can easily interpolate the LDFs in the angular space using an interpolation function of a lower degree. Given a pair of two LDF samples with different incident angles, θ1 and θ2, and material parameters, α1 and α2, a new LDF of incident angle θ and material parameter α can be calculated by the following expression using bilinear interpolation: gθ, α(φ, ψ)= (3) where, / , , / , . PQ ψ n v P(u,v) PO LDF PH φ O u H Figure 3 The relationship between the planar and angular spaces 0 2π φ v u (a) The original planar space π/2 ψ 0 255[level] (b) The corresponding angular space Figure 4 LDFs in two different parameter spaces non-occluded partially occluded occluded bottom search top (level) Figure 5 Determining the diameters of the light beams from mipmap images 3.3. Acceleration of the luminance calculations (1-wθ)(1-wα)gθ1, α1(φ, ψ) + wθ(1-wα)gθ2, α1(φ, ψ) + (1-wθ)wαgθ1, α2(φ, ψ) + wθwαgθ2, α2(φ, ψ), Q (4) Finally the interpolated LDF in the angular space is transformed into that of the planar space by using the relationship between the angular and planar spaces illustrated in Figure 3. Note that we need a reference point for the parameter space conversion. We also use a bilinear interpolation from the reference points corresponding to four LDF samples to obtain a new reference point for the incident angle θ and the material parameter α. We discretize Eq. 2 with a small interval ∆ in u and v directions to calculate the luminance of a surface. Since the size of the interval is related to the diameter of the light beam, the cost of the luminance calculation would be quite high if we use light beams with a small diameter. To address this problem, we determine the diameters of the light beams taking into account the size of the non-occluded area. We first calculate a mask image (an occlusion map) to get the non-occluded areas. The viewpoint is set to the position of the light source and the occluding objects are projected to a screen to generate the mask image. A non-occluded area is corresponding to the pixels of the mask image where no objects are projected. Alternatively, we could make use of a shadow map [13] to determine the non-occluded areas in rendering. If we want to accelerate the rendering process, the mask image helps to determine an appropriate diameter of a beam. 0° 6° 12° 18° 24° 30° LDF sample LDF sample interpolated LDFs (a) Interpolated LDFs (α=0.999) z Incident beam θ y 150 0 x 2 [%] (b) Interpolated LDF (θ=15°) and the relative error distribution 3000 750 150 3000 (c) The simulation model Figure 6 Interpolated LDFs for changing incident angles Next, we make a mipmap [14] of the mask image, and determine the diameters of the light beams. We find the pixels that do not contain any occluded areas at all, traversing the mipmap hierarchy from the top to the bottom as shown in Figure 5. For example, let us assume that white pixels in the mask image are assigned to non-occluded areas and black pixels are occluded areas. We choose only white pixels, not gray or black pixels in each mipmap-level image to get fully nonoccluded areas. The diameters of the light beams are determined by the mipmap-levels in which we first find only white pixels, and the directions of the beams are determined by the positions of the white pixels. The algorithm works well even when the mask image has concave non-occluded areas. 4. EXAMPLES ILLUSTRATING THE PROPOSED METHOD Figure 6 (a) shows interpolated LDFs while changing the incident angles of the light beams. At both ends of the figure are LDF samples generated by a subsurface scattering simulation [6] with a Henyey-Greenstein phase function [15] (g=0.8). We use the simulation model shown in Fig. 6(c). The size of the translucent object is 30003000750 units (1 unit = 1 pixel). The observation area is located at the beam incident point with 150 widths in x, y directions. The incident angle θ is the angle between the incident beam and the normal vector of the simulated object (z direction). We set a reference point at height of 100 units above the brightest point. The brightest point moves to the right in the direction of the incident beam, and the distribution of LDFs is distorted. Figure 6 (b) shows an interpolated LDF (θ=15°) and the relative error distribution of the LDF compared to a LDF generated by the subsurface scattering simulation. Figure 7 (a) shows an image rendered by a method in which the change in incident angles has no effect on scattering. This method is equivalent to most traditional methods. Figure 7 (b) is rendered by the proposed method. Here, light beams of different incident angles and directions are used. The incident angle of the center beam is set to 0° and the outside beams are 45° and 55°. Scattering phenomena depend heavily on the wavelength of light. We use different albedos for the translucent object for the R, G, B color components: αR=0.6, αG=0.9, αB=0.3. We use light beams with seven colors: white (center), red, magenta, blue, cyan, green and yellow. The proposed method uses four LDF samples: (θ1, α1)=(0, 0.2), (θ2, α2)=(0, 0.999), (θ3, α3)=(60, 0.2) and (θ4, α4)=(60, 0.999), where θ and α indicate the incident angles and albedos, respectively. (a) Traditional method (b) The proposed method Figure 7 Subsurface scattering caused by inclined light beams (a) (b) Figure 9 Variable diameters of the beams (incident angle, θ=55°, of a light source located at the left side) (c) (d) Figure 8 Subsurface scattering considering obstacles Comparing the two images, in the proposed method the actual scattering effects described before are realistically rendered, especially the distorted luminance distributions and the shift of the brightest points. The incident points of the light beams are exactly the same in both figures, but the appearances differ because of the shift of the brightest points. The small support of the LDFs causes an artifact at the boundaries of each LDF in Fig. 7 (b), but this can be suppressed by using LDFs with a larger support. As shown in Figure 8, the proposed method is able to render subsurface scattering effects taking into account an obstacle. We use a directional light source and the incident angle of the light is set to 55°. The LDF samples are the same as those of Figure 7. In Figs. 8 (a) and (c), the light comes from the left side, while in Figs. 8 (b) and (d), from the right side. In Figs. 8 (a) and (b), the albedos of the translucent object are set to αR=0.6, αG=0.9 and αB=0.3, while in Figs. 8 (c) and (d), to αR=0.9, αG=0.3 and αB=0.6. The edges between the illuminated and unilluminated areas are blurred and the color of the blurred regions slightly changes because of the scattering property of light depending on the wavelength. Comparing Figs. 8 (a) and (b), the blurred areas are varying depending on the incident directions. Comparing Figs. 8 (a) and (c), the color of the scattered light is varying depending on the material’s albedos. Figure 9 shows an image rendered with variable diameters of beams, while the image in Figure 8 (a) was rendered with a constant diameter of beams. The quality of the images is almost the same. The rendering time was 55.875 seconds for the image in Figure 9, while the method using a constant diameter of the beams (Fig. 8 (a)) required 387.607 seconds with a Dual Core AMD Opteron 2.59 [GHz] CPU and 2 GB memory. The number of light beams was 4107 in Figure 9, and 45600 in Fig. 8 (a). This shows that we are able to accelerate the rendering process without sacrificing the image quality. 5. CONCLUSIONS We proposed a method for rendering a translucent object taking into account subsurface scattering phenomena such as the shift of the brightest point and distorted luminance distributions for the light that slopes in on the surface. The proposed method efficiently calculates the luminance of a surface by convolving Luminance Distribution Functions (LDFs) corresponding to light beams with different diameters, and the LDFs are interpolated from a pair of LDF samples depending on the incident angles of the light beams and the material parameters of the translucent object. The reference point Q is set to a constant height in the proposed method. Better results can be obtained by an extension which would set the reference point based on the luminance distributions of the LDF samples. The method is a promising technique for rendering a variety of translucent materials. Given a few LDF samples, we can get a wide variety of LDFs by introducing an interpolation in multi-dimensional space. For examples, a tri-linear interpolation makes it possible to change two material parameters such as an albedo and a phase function, simultaneously, as well as the incident angle of the light beam. Although in the proposed method presently we assume a planar surface, it is possible to extend the method to be able to handle also curved surfaces with small curvature, such as flat plates. Such an extension can be accomplished with the help of a mapping technique, that is, the surface luminance L can be calculated in the parameter space by which a curved surface is defined, and the obtained luminance mapped onto the curved surface. The proposed method can be further accelerated by utilizing GPU power. 6. REFERENCES [1] H. W. Jensen, S. R. Marschener, M. Levoy, P. Hanrahan: “A practical model for subsurface light transport,” ACM Trans. Graphics 20(3): 511-518 (2001). [2] H. W. Jensen, J. Buhler: “A Rapid Hierarchical Rendering Technique for Translucent Materials,” ACM Trans. Graphics, 21(3): 576-581 (2002). [3] T. Mertens, J. Kautz, P. Bekaert, H. P. Seidel, F. V. Reeth: “Interactive Rendering of Translucent Deformable Objects,” Proc. the 14th Eurographics Workshop on Rendering: 130-140 (2003). [4] Y. Chen, X. Tong, J. Wang, S. Lin, B. Guo, H. Y. Shum: “Shell Texture Functions,” ACM Trans. Graphics, 23(3): 343-353 (2004). [5] C. Donner, H. W. Jensen: “Light Diffusion in Multi-Layered Translucent Materials,” ACM Trans. Graphics, 24(3): 1032-1039 (2005). [6] I. Masuike, T. Tamaki, K. Kaneda: “A Reserch for A Solution Method of Sub-Surface Scattering Equation Using Light Beams,” Proc. IEICE General Conf.: 83 (2007). (in Japanese) [7] K. Takamura, T. Manabe, T. Tamaki, K. Kaneda: “Subsurface scattering simulation and a rendering model taking into account the irradiance distributions,” IEICE Tech. Report, PRMU 109: 37-42 (2009). (in Japanese) [8] A. Furutsuki, K. Takamura, T.Manabe, T. Tamaki, K. Kaneda: “A Rendering Model for Sub-surface Scattering in Translucent Materials based on Irradiance Distributions”, Proc. VC/GCAD Sympo.: 26.1-8 (2011). (in Japanese) [9] M. Shinya, M. Shiraishi, Y. Dobashi, K. Iwasaki, T. Nishita: “Rendering Translucent Materials with Plane-parallel Solution,” J. Information Processing, 17: 180-190 (2009). [10] M. Shinya, Y. Dobashi, M. Shiraishi, K. Iwasaki, T. Nishita: “An Efficient Multiple-Scattering Calculation Method and Its Application to Tree Rendering,” VC/GCAD Sympo.: 18.1-6 (2009). (in Japanese) [11] M. Shinya, M. Shiraishi, Y. Dobashi, K. Iwasaki, T. Nishita: “A Fast Hair Rendering Method with Anisotropic Plane-Parallel Model,” VC/GCAD Symp.: 2.1-6 (2010). (in Japanese) [12] C. Donnor, J. Lawrence, R. Ramamoorthi, T. Hachisuka, H. W. Jensen, S. Nayar: “An Empirical BSSRDF Model”, ACM Trans. Graphics 28(3), Article No.30: 1-10 (2009). [13] L. Williams: “Casting curved shadows on curved surfaces”, Proc. SIGGRAPH’78, 270-274 (1978). [14] L. Williams: “Pyramidal parametrics”, ACM SIGGRAPH Computer Graphics 17(3), 1-11 (1983). [15] L.G. Henyey, J. L. Greenstein: “Diffuse radiation in the galaxy,” Astrophys J. 93: 70-83 (1941).
© Copyright 2026 Paperzz