Docstoc

Reproducing Reflection Properties of Natural Textures onto Real

Document Sample
Reproducing Reflection Properties of Natural Textures onto Real Powered By Docstoc
					N. Kawai: Reproducing reflection properties of natural textures onto real object surfaces. In Texture 2005: Proceedings of the 4th International Workshop on Texture Analysis and Synthesis, pp. 101–106, 2005.

Reproducing Reflection Properties of Natural Textures onto Real Object Surfaces
Naoki Kawai Dai Nippon Printing Co., Ltd. Kawai-N@mail.dnp.co.jp Abstract – Synthetic decorative materials such as synthetic leather and printed wood grain are widely used in our everyday life, but we often notice that they are imitations because they lack richness of the appearance. One of the differences between the appearances of natural materials and synthetic ones is the complicated variance of reflection, but imitations have no means for reproducing the optical effect. In this paper, we first propose a technique for reproducing light and view dependent appearance of modeled CG objects onto real objects by forming anisotropic reflection on real object surfaces using microscopic grooves related with the normal vector field of modeled objects. We also propose an efficient technique to capture the normal vector field of natural materials and it makes it possible to reproduce the appearance of natural textures onto materials like plastics. Our method can be applied on mass production at a low cost, so the method is practically useful to give richness of the appearance on industrial products such as synthetic leather and wallpapers. 1. Introduction complex reflection as images or movies displayed on display devices. In other word, these techniques give natural appearances onto originally silky polygons only in computer graphics. Paying attention to real world, there are many industrial products that imitate natural materials such as synthetic leather and printed wood grain on furnitures or wallpapers. We often feel something is missing with them and finally find that they are imitations. One of the reasons that we are not satisfied with them is that a printed texture shows only a static appearance of the original material that is independent from light or viewing direction. The difference of appearance between natural textures and printed textures is similar to the difference of effect between BTF and conventional texture mapping in computer graphics. The common difference is dynamic behavior of appearance depending on light and viewing direction. Looking at the studies on computer graphics again, an anisotropic reflection model Poulin and Fournier proposed [11] explains that a set of small grooves on the surface of the object causes anisotropic reflection. This model suggests we can move reflection pattern just as we want it to be on real object surfaces if we engraved grooves in carefully determined direction on the surface of the object. Actually, we can see similar effects on some metal engraving works and decorative paper as surface decoration. In some jacquard products and silk fabrics, we can also see the same effect caused by the stream of yarn running on the cloth surface. In the case of these decorative products, the design of stream of grooves is limited to the simple pattern like a concentric circle or hand-drawn pattern that is difficult to amend in the way we want the appearance to be. In this paper, we introduce a technique that generates streams of grooves corresponding to the three-dimensional shape of modeled objects and makes it possible to reproduce dynamic appearance of modeled shape onto the surface of real world object. We also introduce a technique for capturing appearances of natural materials efficiently, which measures material’s normal vector field as one of representative reflection properties. Finally, we make it possible to reproduce dynamic appearance of natural texture onto real objects by combining two techniques described above. The remainder of this paper is organized as follows: Section 2 describes related works and the overview of our approach. Section 3 describes detailed techniques for relating modeled CG objects with grooves and embodying grooves onto real objects. Several results of reproduced appearance are shown. Section 4 describes a fundamental idea and improved techniques for capturing reflection properties of natural texture samples focused on wood grain, and some captured results are shown. Finally, we conclude proposed technique in Section 5 with some discussions.

Various mapping methods for computer graphics have been developed for improving image quality. They are necessary techniques today to produce photo realistic images using limited computational resources without modeling detailed geometry. Texture mapping (or color mapping) is the most representative technique that maps spatially variant albedo captured by scanners or digital cameras. Another representative is bump mapping that maps perturbation to surface normal and makes it possible to simulate light dependent appearance of rough or bumpy surface. Textures and bumps can be defined independently, and the combination of these two techniques is the most common way in computer graphics production as well as traditional diffuse and specular reflection models. Several high-dimensional reflectance functions that describe light and view dependent reflection properties have been also proposed, such as bidirectional reflectance distribution function (BRDF) for homogeneous materials and bidirectional surface scattering reflectance distribution function (BSSRDF) for heterogeneous and translucent materials. They are proposed from the viewpoint of describing light behavior at the first time [9], and their applications for computer graphics are widely discussed [3] [4]. Recently, bidirectional texture function (BTF) has been proposed as a simplified representation by omitting subsurface scattering term from BSSRDF, and its improvement for practical use is one of ongoing studies in computer graphics [1] [8]. These reflection models are used in rendering processes in CG production; therefore we can see the effects of

2.

Strategy and method

2.1 Related work The key idea of this work is to form anisotropic reflection onto real objects corresponding to reflection properties modeled as objects in computer graphics. Here, we briefly take a look at the studies on anisotropic reflection. Poulin and Fournier proposed [11] a model for anisotropic reflection for the purpose of photo realistic rendering. They modeled scratches on object surface as a set of microscopic cylinders, and showed that anisotropic reflection can be rendered by considering shadowing, hiding and inter reflection of cylinders. Though several parameters were introduced for describing reflection properties, the most effective parameter for determining a direction of anisotropy is the direction of cylinders. On the other hand, discussions on anisotropic reflection of several types of textures have been focused on. One of the most typical materials that have complex reflection is cloth. As an example, Xu et. al. [14] proposed a technique for rendering realistic knitwear by introducing basic primitive of knitwear called lumislice. One of another representative materials that show anisotropic reflection is wood grain. Kawai [5] and Marschner et. al. [7] analyzed the fiber structure of internal wood and explained that the roughly flat lumber surface shows anisotropic reflection caused by spatially variant fiber inclination on the surface. Recently, researches on BTF and other high-dimensional reflectance functions are extensively studied and various techniques for capturing light-view-dependent appearance of texture samples are proposed. Not limited for textures but for more general scene, techniques such as light field [6] and lumigraph [2] are proposed as powerful and general-purpose scheme for measuring and reconstructing high-dimensional environment from multiple images. Not only general-purpose techniques for capturing scene or material properties, researches that focused on obtaining necessary properties from limited but often useful categories of objects are also prospering. As a typical research of such an approach, Paris et. al. [10] introduced techniques for capturing properties such as orientation and normal vector of hair from multiple images. 2.2 Outline of method We propose new techniques for reproducing dynamic appearance of natural textures onto real objects from two points of view. One is acquisition of normal vector field from real material sample, and the other is reproduction of anisotropic reflection onto real objects corresponding to given normal vector field. As the first step, we describe a method for making anisotropic reflection onto real objects based on modeled three-dimensional objects in computers. By converting the normal vector field of computer-generated image into the tangent field of scratches on real objects, we form anisotropic reflection that has similar appearance to modeled geometry on the surface of real objects such as plastic film. It can be assumed as a reverse technique of anisotropic reflection model Poulin and Fournier proposed [11] because their model simulates a complex reflection from microscopic structure, while our method forms microscopic structure from reflection properties. We focus on how to control direction of grooves because the most

influencial property to control anisotropic reflection is direction of scratches in Poulin’s model. As the second step, we introduce an efficient capturing method focused on obtaining normal vector field of real texture sample. We explain the method focusing on the anisotropic reflection on wood grain as an example of natural materials. We describe a fundamental idea to begin with, and introduce several improvements, which are mainly on efficiency of the measurement, one by one. 3. Reproducing appearance onto real object

We now describe the entire process for reproducing dynamic appearance of modeled CG objects by forming anisotropic reflection onto real objects. Several results of reproduction are shown with brief discussion. We describe three steps of the procedure in order: converting normal vector field into tangent field, drawing hairline images following generated tangent field and embodying grooves onto real objects. 3.1 Converting normal vector into tangent Our initial conception is to map normal vectors, which are on rendering projection plane, onto the surface of real object one way or another. It means information of a normal vector at each pixel should be used for defining a groove at a corresponding point on the real object. Because a normal vector has two degrees of freedom, which consists of elevation and azimuth around the projection plane, grooves on real objects need to be made reflecting two parameters for complete reproduction of reflection. On the other hand, we can handle only one degree of freedom for controlling the direction of grooves, a tangent. For this difference on dimension, we need to introduce a technique for reducing dimension reasonably. Our purpose is to reproduce a visually similar reflection on real objects, and is not to reproduce complete geometry or complete appearance on it. The reason we use normal vector and tangent for combining CG world and real world is that each of them is the most influential property for controlling light-view-dependent appearance in each of two worlds. We first project a normal vector to each pixel on the projection plane instead of projecting color value in usual rendering process. In a manner of G-buffer Saito and Takahashi proposed [12], the normal vector is to be stored as a content of the buffer, and the G-buffer becomes a normal vector field of modeled scene. In general rendering process, modeled scene or geometry reacts to the lighting condition and projects its intensity onto corresponding position of pixel. From the viewpoint of the projection plane, a normal vector and reflection properties are specified at each pixel, and each of the pixels shows its radiance responding to given lighting condition. It is valuable that the normal vector stored in each pixel keeps the essence of variety of appearance. In other word, a set of response to given lighting of a normal vector and reflection properties stored at each pixel is exactly a rendered image. For the purpose of reducing dimension of normal vector, we introduce a normal projection plane (NPP), a plane that is perpendicular to the projection plane and parallel to a fixed azimuth of α around the projection plane. A normal vector at each position x can be represented by its azimuth and elevation (θx,φx) in the local spherical coordinate system attached to the projection plane, and we project it

perpendicular onto the NPP as illustrated in Figure 1. By this projection, the normal vector is converted into a new vector (α , t ) , and we employ the new elevation t as a tangent value of grooves at a corresponding point x’ on real object. We convert the whole normal vector field into a tangent field by converting the normal vector all over the projection plane. Though the obtained tangent value loses the information about azimuth of the original normal vector, it records projected elevation. That is, it is still possible to reconstruct the angle between lighting direction and the original normal vector if the lighting direction changes along the fixed azimuth α . This means that the scratches placed along the tangent t here can be expected to behave somehow similar to the example geometry. To supplement it intuitively but imperfectly, grooves directed to the same direction respond the same reflection to a given lighting condition as well as surfaces that have the same normal vector and the same reflection property respond the same reflection to a given lighting condition.

paper or plastic film accomplished with heat and pressure. We employ photo-etching process for producing embossing molds. We first prepare a copper plate coated with photoresist, and the binary image of streamlines is exposed onto the plate using laser beam. By developing the resist, the original picture of streamlines appears on the plate and only the exposed streamline area harden. The etching process corrodes only area of unexposed resist (background of streamline image) because the hardened resist keeps copper from corrosion on streamline area, so that streamlines remain as convex lines and the plate becomes a mold for embossing. 3.4 Result Figure 2 shows the sample shape we used for verifying an effect of proposed method. We generated this periodic pattern in procedural way and its height field is shown in the left image. The right image shows converted tangent field. We placed NPP parallel to fixed azimuth α = 0 (horizontally) in this case. Figure 3 (left) shows a part of the image of streamlines generated based on obtained gradient field. Each streamline has been generated with 60 micrometers in width at the center, and rasterized in resolution of 1000 pixels per centimeter. We formed photo-resist from generated streamline image using a gravure plate-making system that exposes a given image onto the copper plate by laser beam. After substantiating grooves onto a mold with 40 micrometers in depth by photo-etching process, we embossed grooves onto a PVC film shown in Figure 3 (right). We can see the shape of the sample geometry on the film with relatively intense specular reflection as the result of spatially variant anisotropic reflection. Figure 4 shows the changes of appearance on the film according to various lighting direction. We can see that the appearance of reflection changes depending on the lighting direction just like the modeled object behaves in three-dimensional environment. Figure 5 shows another result. We used a low frequency random pattern as the original height field. Although the film doesn’t show all the possible appearance as original shape shows in three-dimensional environment, its appearance changes according to (one component of) the lighting direction as we expected in 3.1. At least it has enough effect to let observers recognize three-dimensional shape.

Figure 1: Projection of normal vector 3.2 Drawing hairlines for grooves The tangent field for grooves is obtained in the previous step. As a next step, we draw numbers of streamlines over the field along calculated tangent value at every position. Data visualization for vector fields is widely researched and Turk [13] summarized a strategy for placement of streamlets called hedgehog illustration or vector plots. We employ this approach basically. We first place seeds on a regular grid or on a jittered grid at adequate intervals. Each seed moves a short distance toward the direction of tangent value at the point, and the seed gets a new position with a new tangent there. This operation is repeated until the length of the locus reaches prescribed length. All the points including the initial seed and the final point are stored as control points of a spline curve that represents the locus of the seed. The locus becomes a skeleton of a streamline and we rasterize the locus with a certain width that is the widest at the center and becomes gradually narrower to the tips. Repeating vector plots and rasterizing on all the seeds results in the entire image of streamlines. 3.3 Substantiating grooves onto real object The image of streamlines is just a plan for physical scratches, and we need a means for realizing scratches physically onto objects. From the viewpoint of the cost of mass-production, embossing is one of practical processes that copies three-dimensional form onto materials such as

Figure 2: Sample height field and converted tangent field

Figure 3: A part of the image of streamlines and a physically embossed film

Though the fibers are bent both horizontally and vertically to the lumber surface, we ignore the horizontal inclination for our capturing method. One reason is that the vertical inclination affects much greatly to the appearance than horizontal inclination, and another reason is that we can succeed only one parameter to final grooves, a tangent. 4.2 Estimation of inclination distribution Figure 4: Appearance changing by lighting direction We estimate the inclination γ of fibers at each position by analyzing the response of reflection illuminated from various directions. In practical, it is impossible to illuminate ideal directional light from arbitrary direction, so we use several point lights (or spot lights) as illustrated in Figure 7. In this system, the camera is fixed perpendicular to the sample lumber and captures the appearance of the sample surface with every lighting direction.

Figure 5: Another result 4. Acquisition of texture surface properties

We have introduced a method that gives three-dimensional appearance onto real objects in the previous section. Not only on procedurally generated geometries, our technique can be applied on various representations of geometries in principal because our method requires only normal vector information. In this section, we introduce an efficient method for obtaining necessary normal vector information of real sample texture. The goal of this section is to reproduce the appearance of natural texture onto real objects. We focus on wood grain as an example of natural texture in this section. 4.1 Reflection property peculiar to wood grain The surface of real wood grain shows the peculiar reflection called “figure” that changes its appearance depending on viewing and lighting direction. The flat surface of lumber virtually consists of inclined inner surface of microscopic cylinders [5] [7]. The cylinders are dried fiber cells that have relatively intense specular reflection. As illustrated in Figure 6, the inclination of fibers changes on the lumber surface. The incident light reflects at the inner surface of cylinders toward the direction depending on cylinders’ inclination. The fibers are usually not straight inside wood, but form a certain pattern called “cross grain”. The cross grain causes inhomogeneous distribution of inclination of the fibers on lumber surface. The inclination has similar effect as bump mapping on polygon surface that perturbs normal vector as shown in Figure 6. Thus we employ the inclination γ at each position as tangent of scratches t at the corresponding position.

Figure 7: Capturing apparatus All the light sources, the sample and the camera are fixed and their relative positions are known; therefore the viewing and incident angles can be specified at arbitrary point on the sample. The specular reflection becomes the most intense when the incident direction becomes a direction, which is viewing direction mirrored about normal vector. The inclination γ at a point equals to the angle between the lumber normal and the virtual normal there that is a bisector of viewing and incident direction when the reflection becomes the most intense. For each pixel, we compare all the captured pictures for detecting the brightest picture at the pixel, but the accuracy of γ is depending on the number of lights after all. To reduce the necessary number of pictures, we now introduce the interpolation between limited numbers of incident angle. We can estimate the incident direction β max that gives the most intense reflection as illustrated in Figure 8 by interpolating the captured albedo for each incident angle β . From our experience, six or seven lights are enough to obtain inclination with practical accuracy using B-spline curve for interpolation in most cases.

Figure 6: Reflection caused by fiber cells

Figure 8: Interpolation of incident angle

4.3 Adjustment of illumination We need to analyze albedo of several pixels that are at the same position but in different frame. Here we have the difficulties on giving uniform illumination because we use plural point light sources. It results in the problem that we can’t assume each pixel value as albedo. The problem includes two aspects of uniformities. One is spatial fluctuation of illumination distribution over the sample. The other one is difference of emission properties between light sources that results in different illumination to the same position of different frames. The essence of the problem is unknown illumination at each point in each frame, so we introduce a method for estimating reasonable albedo by adjusting variance of illumination both spatially and between light sources. We take pictures that capture the appearance of a matte white paper illuminated with the same light sources as used for the sample. We assume the white paper as mostly ideal diffuse surface so that we can assume each of the pictures as illumination distribution for each of lighting conditions. We divide pixel value of the sample image by corresponding pixel value of the paper image, and we assume the resulted value as albedo of the sample. Though the assumed albedo is not correct because the picture of the paper doesn’t capture exact illumination distribution, the result is effective enough to correct the inclination of fibers. Figure 9 shows the example of the revision. The images of the sample maple lumber (top row) divided by the images of the white paper (middle row) result in albedo of the sample lumber (bottom row).

cross section and pith of the wood (perpendicular to the ground) and it rarely becomes zero. However, it won’t be an obstacle to shift average in practice because our interest is relative variance of inclination over the lumber surface and preciseness of measurement is not necessary for the purpose of reproducing appearance onto real objects. We take two pictures of the sample texture illuminated from the upper side and the lower side of the sample as shown in Figure 10. We compare two pictures pixel by pixel by subtracting the upper image from the lower image. The results are shown in Figure 11. The left image is generated from the maple sample shown in Figure 10 and the right image is generated from a birch lumber with conspicuous interlocked grain. The absolute value of inclination has been already lost in these figures, and the pixel value indicates only a degree that shows how fibers incline to the upper side.

Figure 10: Albedo illuminated from upper and lower side

Figure 11: Subtracted images (lower - upper) We expect that the wider range of direction of grooves makes the reproduced reflection more attractive. We employ histogram matching for shifting average of inclination to zero and for converting distribution of inclination into wider range. Figure 12 (left) shows the histogram of the subtracted image of the birch sample shown in Figure 11 (right). The histogram shows that the inclination has clustered on one side. Here, we introduce another target histogram as shown in Figure 12 (right) as an ideal distribution of inclination, which is generated from tangent distribution of a sine wave. Each of original pixel value is converted into corresponding pixel value that has the same relative accumulative frequency in target histogram. We finally use distribution of inclination obtained here as a normal vector field for reproducing method we proposed in the previous section.

Figure 9: Albedo adjusted by illumination distribution 4.4 The simplest measurement Even though we simplified the measurement method by introducing interpolation of lighting direction, we still need many pictures for calculating the inclination field. For example, in the case that we use seven light sources, we need seven pictures of the texture sample and another seven of white paper for adjusting illumination. We try to introduce much simpler measurement method by analyzing appearances captured with only two light sources. The goal of this section is to analyze limited number of appearances as best as we can. Here, we introduce a supposition that the average inclination of all the fiber becomes zero. Of course, actual average inclination is depending on the angle between a

Figure 12: Subtracted and target histograms

4.5 Result Figure 13 shows the simulated appearance of measured inclination of maple (top row) and birch (bottom row) samples. In this figure, we applied simple bump mapping on flat surface and calculated only specular reflection using various lighting directions. We can see reconstructed moving reflection of sample materials here.

intensity of reflection by further studies. For example, placing each groove along the tangent direction with certain randomness in direction independently or fizzling streamlines in some extent will decrease intensity of reflection. The proportion of width and depth of grooves will also effect the reflection intensity. References [1] Dana, K. J., Van Ginneken, B., Nayar, S. K. and Koenderink, J. J. 1999 Reflectance and texture of realworld surfaces. In ACM Transactions on Graphics 18, 1 1-34 Gortler, J., Grzeszczuk, R., Szeliski, R and Cohen M. F. 1996. The Lumigraph. In Proceedings of Siggraph 1996, 43-54 Hanrahan, P. and Kruger, W. 1993 Reflection from layered surfaces due to subsurface scattering. 1993. In Proceedings of Siggraph 1993, 165-174 Jensen, H. W., Marschner, S. R., Levoy, M. and Hanrahan, P. 2001. A Practical Model for Subsurface Light Transport. In Proceedings of Siggraph 2001, 511-518 Kawai, N. 1998. Modeling Fiber Stream of Internal Wood. 1998. In ACM SIGGRAPH 98 Conference abstructs and applications, 271 Levoy, M. and Hanrahan, P. 1996. Light Field Rendering. In Proceedings of Siggraph 1996, 31-42 Marschner, S. R., Westin, S. H., Arbee, A. and Moon, J. T. 2005. Measuring and Modeling the Appearance of Finished Wood. In Proceedings of Siggraph 2005, 727-734 Müller, G., Meseth, J. and Klein, R. 2003. Compression and real-time Rendering of Measured BTFs using local PCA. In Vision, Modeling and Visualization 2003, 271-280 Nicodemus, F. E., Richmond, J. C., Hsia, J. J., Ginsberg, I. W. and Limperis, T. 1977. Geometric considerations and nomenclature for reflectance. Monograph 161, National Bureau of Standards (US) Paris, S., Briceno, H. M., and Sillion, F. X. 2004. Capture of Hair Geometry from Multiple Images. In Proceedings of Siggraph 2004, 712-719 Poulin, P. and Fournier, A. 1990. A Model for Anisotropic Reflection. In Proceedings of Siggraph 1990, 273-282 Saito, T. and Takahashi, T. 1990. Comprehensible Rendering of 3-D Shapes. In Proceedings of Siggraph 1990, 197-206 Turk, G. and Banks, D. 1996. Image-Guided Streamline Placement. In Proceedings of Siggraph 1996, 453-460 Xu, Y-Q., Chen, Y., Lin, S., Zhong, H., Wu, E., Guo, B. and Shum, H-Y. 2001. Photorialistic Rendering of Knitwear Using The Lumislice. In Proceedings of Siggraph 2001, 391-398

[2] [3] [4] Figure 13: Simulated reflection of measured inclination Figure 14 shows a result of reproduction. We applied measured inclination of the birch sample on a tangent field of scratches. On this embossed film, we can see the movement of reflection that looks like natural interlocked grain. [5] [6] [7]

[8]

Figure 14: Printed wood grain with embossed scratches

[9]

5.

Conclusion and discussions

[10] [11] [12] [13] [14]

We proposed a method for reproducing three-dimensional appearance of natural textures onto real objects, which consists of two techniques: a simple and efficient technique for capturing reflection property of natural materials and a technique for reproducing three-dimensional appearance of modeled geometries onto real objects. This method makes it possible to give a three-dimensional effect to general plastic products at a low cost and it is suitable for mass-production so that applications on various types of decorative products are promising. Some problems and future works are discussed below. At this time our method generates unexpected regions where the streamlines are spaced too close together or too far apart. This is caused because of the irregularity and convergence of the tangent field. Although Turk and Banks proposed an optimization technique for similar problems [13], their method was designed for relatively low-resolution images. Our method generates image of extremely high resolution, so that much simpler algorithm will be required. The proposed method controls only the direction of scratches depending on the normal vector of original geometry, and all other properties such as the intensity of reflection are ignored. We will be able to reproduce the


				
DOCUMENT INFO
Shared By:
Stats:
views:61
posted:1/16/2010
language:English
pages:6
Description: Reproducing Reflection Properties of Natural Textures onto Real