International Journal of Automation and Computing

04(1), January 2007, 18-24 DOI: 10.1007/s11633-007-0018-3

Image-based Virtual Exhibit and Its Extension to 3D Ming-Min Zhang∗

Zhi-Geng Pan

Li-Feng Ren

Peng Wang

State Key Laboratory of CAD&CG, Zhejiang University, Hangzhou 310027, PRC

Abstract: In this paper we introduce an image-based virtual exhibition system especially for clothing product. It can provide a powerful material substitution function, which is very useful for customization clothing-built. A novel color substitution algorithm and two texture morphing methods are designed to ensure realistic substitution result. To extend it to 3D, we need to do the model reconstruction based on photos. Thus we present an improved method for modeling human body. It deforms a generic model with shape details extracted from pictures to generate a new model. Our method begins with model image generation followed by silhouette extraction and segmentation. Then it builds a mapping between pixels inside every pair of silhouette segments in the model image and in the picture. Our mapping algorithm is based on a slice space representation that conforms to the natural features of human body. Keywords:

1

Virtual exhibition, color substitution, texture morphing, reflection component decomposition, virtual reality.

Introduction

In recent years, E-business has bloomed significantly. Virtual exhibition in E-business can easily overcome the limitations encountered by its traditional counterpart. However, in E-business, a customer can not touch products as he does in traditional exhibition. Therefore, the effect of virtual exhibition is more important. Today’s virtual exhibition systems are often integrated with customization functions, which helps to make rapid responses to customers’ feedbacks and meet various requests respectively. A lot of commercial virtual exhibition systems have been developed during the past decade. Matsushita is a virtual kitchen system, a retail application set up in Japan to help people choose applications and furnishings for a kitchen[1] . In 1997, Mercedes-Benz introduced the “Virtual Car” simulator at the IAA motorshow in Frankfurt, Germany. This simulator allows user to hold a screen in his hand which displays the 3D model of Mercedes-Benz’s new auto and make selection for its color[2] . There are also some special virtual exhibition systems used in culture, education and entertainment areas. The Canadian Museum of Civilization and the National Research Council of Canada collaborated on the production of Inuit3D, one of the six inaugural virtual museum of Canada exhibitions launched in April 2001[3] . In 2004 the Computer Department of Zhejiang University developed a 3D Dunhuang cultural relic exhibition system[4] . However, the rendering effects using traditional computer graphics algorithms are still not realistic enough. Recently, some new techniques were utilized to provide more realistic feelings. For example, ARCO, a cultural relic exhibition system sponsored by the EU IST Framework V programme, used the augmented reality(AR) technique to allow visitors to hold a virtual cultural object in their hands and observe it from different views[5] . Other great limitations of graphics-based exhibition systems include great demands of rendering time and storage, the cost of product modeling Manuscript received September 10, 2006; revised December 11, 2006. This work was supported by 973 Project (No. 2002CB312100) and Key National Natural Science Foundation of China Project on Digital Olympic Museum (No. 60533080), and National 863 High-tech Project (No. 2006AA01Z303). *Corresponding author. E-mail address: [email protected]

and special and expensive equipment required in AR systems. Therefore, only a few huge exhibition projects choose pure graphics-based techniques. To overcome these limitations, many image-based virtual exhibition systems have been developed. These systems use techniques such as image based rendering (IBR). IBR is the combination of image processing, computer graphics and computer vision. It can achieve realistic feelings while avoid the great cost of geometry modeling and rendering. The simplest IBR method is panoramagram, which has many advantages: fast computation, little storage demands and realistic effects. Therefore, it is widely used by little companies. For example, Peace Co. uses panoramic technique to exhibit its furniture products[6] . Panoramagram provides only a few views of products, so it can not give enough flexibility. Another IBR method is to paste image onto a simple 3D model. RNA Cloth Company developed a cloth exhibition system using this method[7] . But these kinds of image-based exhibition systems suffered from lacking powerful customization functions. In this paper, we develop a novel image-based virtual exhibition system. It has all the advantages of image-based systems and provides a powerful customization function: material substitution. Given a photo of a product, our system can show the effects of the product with various materials. First, the boundary of the product should be drawn out. A lot of technologies can be used to do this, for example, snapping and matting[8−10] . The accuracy of boundary will not affect the substitution results very much so we do not discuss it in this paper. During substitution, we need to generate the geometry of the product, which is the key point in IBR. Several methods are able to perform user-guided modeling from a single image. Criminisi et al.[11] provided a very accurate but the most labor-intensive approach by using projective geometry constraints[12] to calculate 3D locations of user-specified points. Zhang et al.[13] modeled free-form scenes using constraints placed by users, then optimized for the best 3D model to fit these constraints. In 1997 “Tour in image” was brought forward by Horry et al.[14] which created a “spidery mesh” interface to allow the user to specify the coordinates of foreground objects and the vanishing point, then tour

M. M. Zhang et al./ Image-based Virtual Exhibit and Its Extension to 3D

in the image. Inspired by [14], a fully automatic method was provided by Hoiem et al. to build a pop-up model from one image[15] . This method divided the image into coarse categories then “cut and fold” the image. For our system, [11,13] are too labor-intensive for real application, and models generated by [14,15] are too rough to reflect products shapes. In our system, we provide two simple texture morphing methods to generate 3D feelings of products with user-guides. A few simple interactions should be manipulated by users, which are very natural for a human to learn and use. The interaction course is WYSWYG, which allows users to make a balance between exhibiting effects and labor-intensity. Color substitution is another core algorithm in our virtual exhibition system. Not only the result color should be reasonable but also the brightness distribution on the object surface should be preserved very well. Recently a lot of color substitution algorithms were put forward. Reinhard et al. provided a method to transfer the color of source image onto the target image [16] but it could only get satisfied effects when the source and target images contained similar scenes. Inspired by [16], Welsh et al. designed a method to colorize a grayscale image[17] . Some rectangles were divided manually from the source and target images, then a matching process was manipulated for every couple of rectangles. The division result would affect the colorized result significantly. In 2004, Levin et al. brought forward a new colorizing method. User could use color strokes to set new colors for each area to be colorized on the image[18] . All of these algorithms kept the brightness channel of the target image unchanged while using new chromaticity to build new colors. Actually, it is not true according optics theory. We utilize a physical based color substitution algorithm in our system. Comparing with traditional methods, our algorithm can not only preserve the brightness distribution but also ensure the physical color accuracy after substitution. The rendering speed of our system is very fast, and its demand of storage are very little too. Therefore, it is very suitable for Internet. To support the 3D virtual exhibit, we need to reconstruct 3D models of human being. Recently, there have been increasing interests in modeling articulated human body for researchers in virtual environment, computer games, and digital entertainment. People like to meet each other in real life, they expect to be able to do so in virtual environment, too. So building a model for the virtual human is a prerequisite. The model should have realistic appearance and easy to be animated. However, building an animatable human model is not an easy process. Difficulties lie in many aspects. The shape of a human body is rather complex but the vision system is highly developed to distinguish small error in bodies. Furthermore, the body is an elastic organ. Thus a good model must deform according to the surface inflection of the human bodies when people move. Furthermore, because of the rapid progress of virtual environments, games and entertainments, the demand for human body keeps on increasing steadily. Existing modeling method can not fulfill this task. They are either too expensive to deploy or too complicated to learn, and thus can not be deployed in large scale. All these factors call for easy

19

and low-cost modeling methods. A variety of body modeling methods exist[19] . They fall into 4 major categories: authoring, capturing, interpolating, and model fitting. For authoring methods, authors create a model from scratch by interacting with low-level building primitives such as vertices and edges. They often resort to commercial modeling software such as Maya for this purpose. Although these methods give maximal control to the author, they require high artistic talents and good proficiency of the underlying software. Recently, capturing methods became more popular. They use special equipment ranged from stereos[20] to 3D scanners. The resulting model often has holes in it and is over complex. More importantly, it must go through a hard and tedious rigging procedure to be animated. Then there are data-driven methods, which generate new model by interpolating[21] or segmenting then re-compositing[22] existing sets of example models. The last category is model-fitting methods[23,24] , which derive a new model by using information extracted from pictures to modify the generic one. The resulting model is visually appealing and animatable. The methods in the last category best suit our need. There are two typical kinds of model fitting methods. The first one presented by Hilton et al.[23] uses four orthogonal pictures to capture shape details of an individual. The core of the method was to project the generic model to the image space, make modification in 2D, then project back to 3D. We propose an easy and efficient method for building articulated human body. No expensive hardware and software is required. And the modeling process is intuitive. Taking a generic model and two pictures of an individual as input, the method extracts shape features shown in the pictures and embeds them into the generic model.

2

Color substitution algorithm

In our virtual exhibition system, users are allowed to select the materials of a product freely and check the substitution effects. Using this function the user can decide which kind of material is the most suitable for him. It can also help providers collect user feedbacks quickly and exactly, which are very important for customization. First of all, an original picture of product is captured. Then the photo will be analyzed by the analysis module to generate some character information, such as surface reflection components and 3D shape coordinates. In this stage, some interactions should be provided in WYSWYG manner. The analysis module is manipulated only once. Therefore, the rendering time of each substitution process is very fast. Users can freely select different materials for a product and check results immediately. In our virtual exhibition system we designed a novel color substitution algorithm based on the dichromatic reflection model to separate the object color and scene illumination[25] . The object color was subsequently substituted while keeping the illumination distribution unchanged. A new method was employed to estimate the object color. We also developed a set of new parameters to adjust the intensity distribution on the resultant image. Our algorithm is totally automatic and can achieve more re-

20

International Journal of Automation and Computing 04(1), January 2007

alistic effects compared with other color substitution methods. One limitation of our algorithm is that it can only handle objects with pure color. Because it is very hard to decide whether the discontinuation of illumination distribution is caused by lighting conditions or surface colors automatically.

2.1

Dichromatic reflection model (DRM)

Surface reflection is a very complicated physical optics course. There are a lot of reflection models presented by researchers to simulate this course. Most of them are too complex to be calculated fast. In 1985 Shafer introduced a simplified model for inhomogeneous substances, which is named as Dichromatic Reflection Model (DRM)[26] . The most important assumption of this model is NIR assuming that the spectral composition of interface reflection component is approximately the same as that of incident light, except for intensity change. The core equation in DRM is shown as follows: Y (θ, λ) = cd (θ)Sd (λ)E(λ) + Cs (θ)E(λ)

(1)

where Y is the spectral power distribution of reflected light, λ is the wavelength parameter, θ represents the reflection geometry including incident angle, viewing angle and phase angle. Sd (λ) is the diffuse spectral surface reflectance function. E(λ) is the spectral power distribution of the incident light. cd (θ) and Cs (θ) are the geometric factors of the diffuse and specular parts. For all the color signals in the object area, (1) can be rewritten as follows: Cp = αP Cb + βp Ci

(2)

where the subscript p means a pixel, Cp is the pixel’s color, Cb , Ci are the object color and illumination color, respectively. αp , βp reflect the geometric factors. That means every color signal in the object area is the linear combination of two color vectors. For images recorded by a camera, the sensor outputs of R, G and B channels are balanced to remove the effects of the color temperature of light source, which is called white balance. White balance is achieved by adjusting the device sensitivity functions to satisfy Z Z Z E(λ)RR (λ) dλ = E(λ)RG (λ) dλ = E(λ)RB (λ) dλ w

w

however, not all the samples are on Φ exactly. In other words, not all the projected dots on Ψ are exactly on the same line, due to the non-linearity of the camera output and the information loss caused by image compression and other manipulations. In this case, an adjust process is employed, in which the average slope of the lines from the origin to these projected dots is used to decide the line which fits them best. Then the normalized object color vector Cbn can be calculated.

2.3

Color substitution

First we decompose all the input signals with these two vectors in a least square sense to calculate the parameter sets αS and βS using the normalized object and illuminant colors. In the traditional DRM algorithms, the maximum value in the set αS , αmax , is thought to be the amplitude of Cb [27] , which is corresponding to a place where the incident angle is very small and the including angle between the view and reflection directions is very large. Unfortunately, such a place does not necessarily exist in normal scenes. In most cases, we can not estimate the amplitude of Cb accurately. Instead of calculating the exact amplitude of the object color, which is almost impossible in most cases, we select a special dot on the α − β plane, whose color is called the reference color. We use the reference color as the actual color to be substituted. We first select all the dots whose β values are near zero. Then choose the maximum α value among them as the amplitude of the reference color. To keep the intensity contrast scope proportional to the new color, a new parameter format should be developed. As Fig. 1 shows, XB is color B’s projection along the reference color’s direction and it represents B’s intensity on the object color component. Therefore, the new parameter form is given as 8 α + β cos(θ) < 0 α = (4) n : β0 = β where n is the amplitude of reference color. In the substitution course, the parameters α0 and β 0 are calculated for each color signal, and new color is calculated as follows: Cp0 = α0 Cn + β 0 Γ

(5)

where Cn is the new color.

w

(3) where R represents the sensitivity function. From (2) and (3)˘ we find that ‹√can ‹√ ‹√ the ¯ normalized illumination color Cin is 1 3,1 3,1 3 for the images under white balance.

2.2

Normalized body color estimation

As dichromatic reflection model describes, all color signals in the object area should be on the plane Φ spanned by vectors Cb and Ci . When they are projected onto the plane Ψ , which is perpendicular to Φ and across the original point, all of the resultant dots are on a line section PO on Ψ . The vector corresponding to dot P has the same direction with Cb , therefore we can calculate the coordinates of P and project it back to find the vector Cb . We can simply select the farthest vertex (according to the original point) as the projected dot of Cb . In practice,

Fig. 1

2.4

New parameters calculation

Experimental results

Here we give some substitution results using our algorithm. Fig. 2 shows the result for the surface which has

21

M. M. Zhang et al./ Image-based Virtual Exhibit and Its Extension to 3D

subtransparent characteristic. Fig. 3 is an example of broad dynamic brightness range. For many textile products,

Fig. 2

grid, as Fig. 4 shows. The horizontal frame lines are green, and the vertical ones are blue. These frame lines are cut or expended to construct a close texture region. Then some necessary preprocessing should be performed.

Color substitution result 1 Fig. 4

Texture frame lines

We use four outmost frame lines to construct an enclave polygon of the gird. Then other frame lines are cut or expended to reach the polygon.

3.2

Fig. 3

Color substitution result 2

patterns will cause problems during color substitution. As described in previous sections, discontinuations of illumination distribution caused by patterns are confused with that caused by lighting conditions, for example shadows.

3

Grid-based texture morphing

When we exhibit textile products such as garments or curtains, the gauffers and folds must be preserved after material substitution. Gauffers and folds affect not only brightness distribution but also texture directions. The color substitution algorithm introduced in Section 2 will ensure the brightness distribution unchanged. But it is almost impossible to simulate a surface with many gauffers and folds by sub-planes. In this section we will introduce a grid-based texture morphing method[28] , by which users can define a texture grid interactively. Then for each pixel in the object area, the texture coordinates are calculated by means of this texture grid. This method consists of four steps: Step 1. Input and preprocessing. Step 2. Generation of texture gird. Step 3. Local adjustment of texture grid. Step 4. Mapping texture image onto target area.

3.1

Input and preprocessing

All the inputs of this method are horizontal and vertical texture frame lines, which identified the shape of the texture

Generating virtual texture grids

The texture grid is built based on the processed frame lines. First, we calculate the gradual frame lines based on the given horizontal and vertical frame lines respectively. These lines should change gradually from one given line to another to construct virtual horizontal and vertical texture grids. To calculate the gradual frame lines by interpolation, the texture frame lines must be parameterized. Let Li denote one gradual frame line, the discrete parametric representation of Li can be defined as ( Li =

Xi(t) . Y i(t)

(6)

For each vertex in line Li , there is a corresponding parameter t, which is defined as the ratio of the distance from the beginning vertex to the current one and the total length of the line. For Li , a set of t is calculated. We calculate the union of the parameter sets of these input lines and resample these lines using this union. Finally, all the given texture frame lines are discretely parameterized based on the same parameter set. As the methods of generating these two grids are similar, we only discuss the horizontal one. We use linear weighted interpolation to calculate the gradual lines. Given a line Li , the y coordinate of the center point of its bounding box is denoted as the weighting factor wi . Using the weighting wgj , the gradual line’s parametric representation is given as n P

Xj(t) =

i=1 n P i=1

n P

Xi(t) |wi−wgj|

, Y j(t) = 1 |wi−wgj|

i=1 n P i=1

Y i(t) |wi−wgj|

(7) 1 |wi−wgj|

where Xi (t) and Yi (t) are the parametric representations of input frame line Li , and n is the number of the input frame lines.

22

3.3

International Journal of Automation and Computing 04(1), January 2007

Blending

After the generation of the two virtual horizontal and vertical texture grids, the final texture grid can be simply built by their weighted combination. Usually, users pay similar attention to the horizontal and vertical texture directions. Therefore, we use the same weights for the two virtual grids. In some cases, however, users may pay more attention to one texture direction than the other, and the characteristics of the final grid should reflect this accordingly. For example, in the simulation of a sleeve, the vertical frame lines will affect the final grid more than the horizontal ones. In this case we should assign suitable weights for these two virtual grids. Supposing we have built a grid using the same weights, this grid will reflect the characteristics of the two virtual grids equally. We take this grid as the reference grid. The differences between the reference grid and the two virtual grids are calculated. These differences should indicate whether the weights adjustment step is necessary. If the differences are big enough, new weights are calculated according to the bending degrees of the two kinds of texture frame lines. Fig. 5 shows the adjusted result.

Fig. 5

4

Fig. 6

4.1

System flowchart

Generic model building

Our method uses a generic model composed of two parts: a joint tree structure and set of part geometry. Our generic model conforms to H-Anim[29] LOA1 joint hierarchy (as in Fig. 7 (b)). It is better than the one (as in Fig. 7 (a)) in Hilton’s method[23] . Hilton used an over simplified segmentation scheme demanded by his mapping algorithm. The difference is shown in Fig. 7. Notably, in his scheme the upper torso and a large part of upper arm are in a single segment, which is problematic and can cause breaks and/or overlaps when the model is animated.

The weight adjustment result

3D human model based on images

reconstruction

Our method expects two inputs: 1) the pictures (referred to as captured images in the context) of an individual taken from the front and side views, and 2) a generic human model. The method first calculates the camera setting for capturing those images, then renders two images (referred to as model image in the context) with this setting. Next, body silhouettes in all images are extracted and segmented into correspondent parts. The core of our method is an improved two-dimensional mapping algorithm based on cut-line space representation, which builds correspondence between points inside every pair of correspondent parts. After building this twodimensional mapping, the method projects every point in the generic model to image space, computes its corresponding pixel in captured image by applying the above 2D mapping, and finally inversely projects the new pixel to 3D space to get the new 3D position of that point in the generic model. Images from front view modify the width and height of model whereas those from left view modify the depth and height. The final result model is a combination of these two modifications. Fig. 6 shows the flowchart.

Fig. 7 Model segmentation: (a) Hilton’s scheme[23] : upper torso and large part of upper arm are mixed into a single segment; (b) our scheme: the border between torso and arm is slant, which is more natural

4.2

Image generating and segmenting

In the second part of preparation stage, camera settings of captured images were calculated. Since these pictures are known to be taken from exact front and side views, only the focus length and the distance between the camera and the person need to be calculated. Then simulated pictures (model images) are generated under the same camera settings. With all these images, silhouettes are extracted and segmented into parts. Output of this stage is four sets of borders of all body parts: front model borders, side model borders, front-captured borders, and side-captured borders.

4.3

2D-mapping

This section describes our new 2D-mapping algorithm based on cut-line space deformation. Given the model border and captured border produced in the previous section, the 2D-mapping algorithm will map every point inside the model border to a point inside the captured border. The simplest case is shown in Fig. 8. Every border is

23

M. M. Zhang et al./ Image-based Virtual Exhibit and Its Extension to 3D

composed of left and right halves. The line (referred to as start cut-line) from the start point of left side to the start point of the right side is parallel to the line (referred to as end cut-line) from the end point of left side to the end point of right side. For every point inside the border, there exists a line which passes through it and is parallel to the top line. The union of these lines will cover the whole space inside the border. Such lines are referred to as cut-line. Therefore, a point inside the border can be defined by two parameters, u, which describes the relative position of the underlining cut-line in the set of cut-lines, and v, which describes the relative position of the point on the cut-line. (a)

Fig. 8

4.4

2D mapping: coordinates system mapping

3D location modifying

Every 3D point (X, Y, Z) in the generic model is projected to a 2D point (x, y) with a depth value z. The 2D mapping algorithm will then find the corresponding 2D 0 0 point (x , y ) in the captured image. Combining this new 2D location with the depth value obtained before forms a triad, which can be inversely projected to a 3D point. After changing the positions of all 3D points, the model will exhibit similar shape to that in the captured image. Modification based on pictures of a single view only changes the shape orthogonal to the viewing direction and is incomplete. Full adjustment is the combination of the modifications obtained from two views. Repeating this procedure for every point in the generic model will give us the new model. The above algorithm gives the experimental results shown in Fig. 9. The generic model is captured by 3D whole body scanners. We fill the hole with the method in Blender, and cut it into segments with a modified version of the builtin “knife” script. Besides the obvious resemblance of height and weight between the result model and the picture, local features such as delicate contour details in arms and legs are also captured. Neither seams nor overlaps exist. And because of the structure inherited from the generic model, the result is ready to be animated. Neither does it require the pose in the picture to be similar to that in the generic model, which helps the potentially wide use of our method.

(d)

(b)

(e)

(c)

(f)

Fig. 9 Results. (a) generic model. (b), (c), and (d) result models from front, side, and new views. (e) and (f) silhouettes extracted from front and side pictures

5

Conclusion

In this paper we discuss the techniques for virtual exhibit, color substituting algorithm and block-based texture mapping method. And the system is extended to 3D environment. An improved method for building articulated human body, which deforms the generic model with features extracted in individual pictures, is presented. Our contribution is the novel slice space representation conforming to the natural feature of human body, which captures more details in individual pictures and preserves the connectivity of surface in more cases.

References [1] Matsushita. Wired 2.07: Kitchen Magician, [Online], Available: http://www.wired.com/wired/archive/2.07/kitchen.html, 8 Dec. 2006. [2] Benz Mercedes. Mercedes-Benz Website, [Online], Available: http: //www. mercedesbenz. com/e/home/iaa/iaa pr10. html, 8 Dec. 2006. [3] F. Corcoran, J. Demaine, M. Picard. INUIT3D: An Interactive Virtual 3D Web Exhibition. In Proceedings of the Conference on Museums and the Web, Boston, Massachussets, pp. 18–20, 2002. [4] Y. Liu, D. M. Lu, C. Y. Diao, Y. Zhu. Dunhuang 285 Cave Multimedia Integrated Virtual Exhibit. Journal of ComputerAided Design & Computer Graphics, vol. 16, no. 11, pp. 1528– 1534, 2004.

24

International Journal of Automation and Computing 04(1), January 2007

[5] M. White et al. ARCO-An Architecture for Digitization, Management and Presentation of Virtual Exhibitions. In Proceedings of International Conference on Computer Graphics, Hersonissos, Crete, pp. 622–625, 2004. [6] Peace Home Inc. Peace Home, [Online], Available: http://www.jj5u.com/hwm/company.asp, 8 Dec. 2006. [7] RNA2006. RNA. Cloth Exhibition System, [Online], Available: http://rnainc.jp/index2.html, 8 Dec. 2006. [8] Y. Li, J. Sun, C. K. Tang, H. Y. Shum, Lazy Snapping. ACM Transactions on Graphics, vol. 23, no. 3, pp. 303–308, 2004. [9] C. Rother, V. Kolmogorov, A. Blake. “GrabCut”: Interactive Foreground Extraction Using Iterated Graph Cuts. ACM Transactions on Graphics, vol. 23, no. 3, pp. 309–314, 2004. [10] J. Sun, J. Jia, C. K. Tang, H. Y. Shum, Poisson Matting. ACM Transactions on Graphics, vol. 23, no. 3, pp. 315–321, 2004. [11] A. Criminisi, I. Reid, A. Zisserman, Single View Metrology. International Journal of Computer Vision, vol. 40, no. 2, pp. 123– 148, 2000. [12] R. Hartley, A. Zisserman, Multiple View Geometry in Computer Vision, Cambridge University Press, Cambridge, UK, 2004. [13] L. Zhang, G. D. Phocion, J. S. Samson, S. M. Seitz. Single View Modeling of Free-form Scenes. In Proceedings of the IEEE International Conference on Computer Vision and Pattern Recognition, IEEE Computer Society, Los Alamitos, CA, USA, vol. 1, pp. 990–997, 2001. [14] Y. Horry, K. I. Anjyo, K. Arai. Tour into the Picture: Using a Spidery Mesh Interface to Make Animation from a Single Image. In Proceedings of the 24th Annual Conference on Computer Graphics and Interactive Techniques, ACM Press/AddisonWesley Publishing Co., New York, NY, USA, pp. 225–232, 1997. [15] D. Hoiem, A. A. Efros, M. Hebert. Automatic Photo Pop-up. In Proceedings of ACM SIGGRAPH 2005, ACM Press, New York, NY, USA, vol. 24, no. 3, pp. 577–584, 2005. [16] E. Reinhard, M. Adhikhmin, B. Gooch, P. Shirley. Color Transfer between Images. IEEE Computer Graphics and Applications, vol. 21, no. 5, pp. 34–41, 2001. [17] T. Welsh, M. Ashikhmin, K. Mueller. Transferring Colour to Grayscale Images. In Proceedings of ACM SIGGRAPH 2002, San Antonio, Texas, pp. 277–280, 2002. [18] A. Levin, D. Lischinski, Y. Weiss. Colorization Using Optimization. ACM Transactions on Graphics, vol. 23, no. 3, pp. 689–694, 2004. [19] N. Magnenat-Thalmann, H. Seo, F. Cordier. Automatic Modeling of Virtual Humans and Body Clothing. Journal of Computer Science and Technology, vol. 19, no. 5, pp. 575–584, Sep. 2004. [20] F. Devernay, O. D. Faugeras. Computing Differential Properties of 3-D Shapes from Stereoscopic Images without 3-D Models. In Proceedings of the IEEE International Conference on Computer Vision and Pattern Recognition, Seattle, WA, pp. 208–213, 1994. [21] P. P. Sloan, C. Rose, M. Cohen. Shape by Example. In Proceedings of the 2001 Symposium on Interactive 3D Graphics, ACM Press, New York, NY, USA, pp. 135–143, 2001. [22] T. Funkhouser et al. Modeling by Example. ACM Transactions on Graphics, vol. 23, no. 3, pp. 652–663, 2004. [23] A. Hilton, D. Beresford, T. Gentils, R. Smith. Virtual People: Capturing Human Models to Populate Virtual Worlds. In Proceedings of IEEE International Conference on Computer Animation, Geneva, Switzerland, pp. 174–185, 1999. [24] W. S. Lee, J. Gu, N. Magnenat-Thalmann. Generating Animatable 3D Virtual Humans from Photographs, Computer Graphics Forum, vol. 19, no. 3, pp. 1–10, 2000. [25] Z. G. Pan, P. Wang, J. H. Xin, M. M. Zhang, H. L. Shen. An Illumination Distribution Preserved Colour Substitution Algorithm Based on Dichromatic Reflection Model. Displays, vol. 26, no. 3, pp. 121–127, 2005. [26] S. Shafer. Using Color to Separatere Reflection Components. Color Research and Applications, vol. 10, pp. 210–218, 1985. [27] S. Tominaga, N. Tanaka. Estimating Reflection Parameters from a Single Color Image. IEEE Computer Graphics and Applications, vol. 20, no. 5, pp. 58–66, 2000.

[28] P. Wang, M. M. Zhang, J. H. Xin, Z. G. Pan, H. L. Shen. An Image-based Texture Mapping Technique for Apparel Products Exhibition and Interior Design. Displays, vol. 24, no. 4–5, pp. 179–186, 2003. [29] C. Babski et al. Humanoid Animation Specification, [Online], Available: http://www.h-anim.org/Specifications/HAnim200x/ISO IEC FCD 19774/, 8 Dec. 2006.

Ming-Min Zhang received her B. Sc. and M. Sc. degrees from Computer Science Dept, Nanjing University in 1990 and from Zhengjiang University in 1995, respectively. She is currently an associate professor of Computer and Engineering Department, Zhejiang University. She has published more than 20 papers in recent years. She is the co-author of two books related to computer graphics and multimedia. Her research interests include virtual reality/virtual environment, multi-resolution modeling, real-time rendering, distributed VR, visualization, multimedia and image processing. Zhi-Geng Pan received his B. Sc. and M. Sc. degrees from Computer Science Dept, Nanjing University in 1987 and 1990, respectively, and the Ph. D. degree from Computer Science and Engineering Department, Zhejiang University in 1993. He is currently a professor of Computer and Engineering Department, Zhejiang University. He has published more than 70 papers in recent years. He is the co-author of two books related to computer graphics and multimedia. His research interests include virtual reality/virtual environment, multi-resolution modeling, real-time rendering, distributed VR, visualization, multimedia and image processing. Li-Feng Ren received his B. Sc. degree in engineering from Hebei University of Technology in 1996. Currently, he is a Ph. D. student in the Department of Computer Science at Zhejiang University, China. His research interests inlcude human body modeling, character animation, and real-time rendering.

Peng Wang received his Ph. D. degree in engineering from Computer Science Department, Zhejiang University in 2005. His research interests include virtual reality/virtual environment, distributed VR, visualization, multimedia and image processing.