Optimized Scalable Wavelet-Based Codec Designs for Semi-Regular 3D Meshes

3D graphics applications make use of polygonal 3D meshes for object’s shape representation. The recent introduction of high-performance laser scanners and fast microcomputer systems gave rise to high-definition graphics applications. In such applications, objects with complex textures are represented using dense 3D meshes which consist of hundreds of thousands of vertices. Due to their enormous data size, such highlydetailed 3D meshes are rather intricate to store, costly to transmit via bandwidth-limited transmission media, and hard to display on end-user terminals with diverse display capabilities. Scalable compression, wherein the source representation can be adapted to the users' requests, available bandwidth and computational capabilities, is thus of paramount importance in order to make efficient use of the available resources to process, store and transmit high-resolution meshes.


Introduction
3D graphics applications make use of polygonal 3D meshes for object's shape representation. The recent introduction of high-performance laser scanners and fast microcomputer systems gave rise to high-definition graphics applications. In such applications, objects with complex textures are represented using dense 3D meshes which consist of hundreds of thousands of vertices. Due to their enormous data size, such highlydetailed 3D meshes are rather intricate to store, costly to transmit via bandwidth-limited transmission media, and hard to display on end-user terminals with diverse display capabilities. Scalable compression, wherein the source representation can be adapted to the users' requests, available bandwidth and computational capabilities, is thus of paramount importance in order to make efficient use of the available resources to process, store and transmit high-resolution meshes.
State-of-the-art scalable mesh compression systems can be divided into two main categories. A first category includes codecs that directly compress the irregular topology meshes in the spatial domain. In such codecs, the connectivity information is encoded losslessly while mesh simplification methods such as vertex coalescing (Rossignac & Borrel, 1993), edge decimation (Soucy & Laurendeau, 1996) and edge collapsing (Ronfard & Rossignac, 1996) are employed to encode geometry. These mesh simplification methods progressively remove those mesh vertices which yield the smallest distortion. In order to enable the reconstruction of the original mesh at various levels of detail (LODs), the discarded vertices are encoded in the compressed bit-stream. Mesh compression systems belonging to this category include Progressive Meshes (Li & Kuo, 1998), (Pajarola & Rossignac, 2000) and Topological Surgery (Taubin et al., 1998). These techniques generally exhibit two major drawbacks: first, due to the highly irregular topology of the input mesh, a large source rate is needed for lossless encoding of connectivity. Secondly, encoding the removed vertices in the compressed bit-stream is quite costly for highresolution meshes. Therefore, such schemes are not useful for complex meshes containing a large number of vertices. An alternative that solves the problem of the large source rates needed to encode the connectivity information, described above, is remeshing, which can be used to convert the original irregular mesh into a mesh consisting of regular elements, such as B-spline (Eck & Hoppe, 1996) or subdivision connectivity patches (Eck et al., 1995). The regular mesh lends itself better to compression, and hence compared to the irregular mesh a much lower rate is needed to losslessly encode its connectivity information. Furthermore, multiresolution techniques alleviate the second problem of having to encode all the original vertices, because only detail information has to be encoded in order to create multiple LODs (or multiple resolution levels). Remeshing together with subdivision-based multiresolution (Lounsbery et al., 1997) are the two major components of the second category of codecs which use space-frequency dilation methods such as wavelet transforms to decorrelate the input mesh data (Khodakovsky et al., 2000), (Denis et al., 2010b). The generated wavelet coefficients are compressed using tree-based bit-plane coding methods (Shapiro, 1993), (Munteanu et al., 1999b) to achieve high compression efficiency. Multiresolution mesh compression techniques provide substantial compression gains compared to their competing schemes, and in this chapter we will confine our discussion to these techniques only.
In the recent past, several multiresolution scalable mesh compression schemes have been proposed. The majority of these schemes use coding techniques which were specifically developed for image compression. However, in general, image and mesh data exhibit different statistical characteristics as the images are consisting of pixels (with intensities) while mesh data involve geometry, i.e., the positions of vertices in a 3D space. Thus, one must be cautious when extrapolating image compression techniques towards mesh geometry encoding.
In this book chapter, we propose a constructive design methodology for multiresolutionscalable mesh compression systems. The input mesh is assumed to possess subdivision connectivity (Lounsbery et al., 1997), i.e., the connectivity in the mesh is built through subdivision 1 . A 3D mesh with subdivision connectivity is also referred to as a semi-regular mesh. With respect to design, we address two major aspects of scalable wavelet-based mesh compression systems, namely, the optimality of embedded quantization in scalable mesh coding and the type of coefficient dependencies that can assure the best compression performance. In this context, thorough analyses investigating the aforementioned aspects are carried out to establish the most appropriate design choices. Later on, the derived design choices are integrated as components of the scalable mesh coding system to achieve state-ofthe-art compression performance.
The remainder of the book chapter is organized as follows: in Section 2, a brief overview of multiresolution analysis of the mesh geometry is given. Section 3 presents a model-based theoretical investigation of optimal embedded quantization in wavelet-based mesh coding. An information theoretic analysis of the statistical dependencies among wavelet coefficients and the conclusions regarding the best exploitable statistical dependency are detailed in Section 4. Section 5 gives an overview of the state-of-the-art mesh compression systems.

Multiresolution analysis of semi-regular meshes
is generally represented as a set of two components, a vertex list c and a polygon list p . c is a matrix whose ith row i c contains the x , y and z position of the ith vertex, i.e.,      ,,, ,, ii x i y i z cc c c. p is a list of polygons made up of edges where each edge is a line connecting two vertices. In computer graphics, 3D meshes are constructed using different polygonal shapes, e.g., triangles, rectangles etc. However, in this chapter, we will confine our discussion to the triangular meshes only.
In the following, a brief theoretical overview of semi-regular multiresolution analysis is presented. Later on, two practical transforms, namely the lifting-based wavelet transform and the spatially adapted wavelet transform are detailed.

Subdivision surfaces
Subdivision is a process of iteratively refining a control polyhedron 0 M into fine geometry polyhedra such that the refined polyhedra 123 , , ...

MMM
converge to a limit surface  M . In general, subdivision schemes consist of splitting and averaging steps. In the splitting step, each triangular face is split into four sub-triangles by adding new vertices. This way, an intermediate polyhedron  A commonly used subdivision is Butterfly subdivision (Dyn et al., 1990). The subdivision stencil for Butterfly is shown in Fig. 1, where the position of a newly introduced vertex p is computed as, ii i ca c whereby i a 's denote the Butterfly weights (Dyn et al., 1990). Loop (Loop et al., 2009) and Catmull-Clark (Catmull & Clark, 1978) are among the other commonly used subdivision schemes for 3D meshes.

Multiresolution analysis
Lounsbery (Lounsbery et al., 1997) first invented the multiresolution analysis for arbitrary topology semi-regular surfaces using subdivision. He proved that refinable bases exist when a coarse mesh 0 M is refined through subdivision, i.e., , for x  0 M and   0 j J . (1)  j x  in the above equation denotes the row vector of scaling functions  j i . Given these refinable scaling functions, scalar-valued function spaces associated with the coarsest geometry 0 M are defined as (Lounsbery et al., 1997): Eq (1) implies that these spaces are indeed nested, i.e., The wavelet space is defined as a space which is the orthogonal complement of together can represent any scalarvalued piecewise function in the space is a row vector containing refinable bases functions of   0 j WM, the following stands (Lounsbery et al., 1997): Combining (1) by combining Eq (5) and Eq (6) From Eq (8), Eq (9) one derives the forward wavelet transform, given by: www.intechopen.com  (Lounsbery et al., 1997).
where d j is a matrix containing the wavelet coefficients for the th j level of the transform.
In general, after the transform, a fair amount of correlation still exist between x , y and z wavelet component. Local frame representation (Khodakovsky et al., 2000) of wavelet coefficients is often used to make wavelet components much more independent. After the local frame transformation, each wavelet coefficient consists of a normal and two tangential components.
s A j and s B j are matrices representing the low and the high-pass filters, respectively, also referred to as analysis filter pairs.
A similar reasoning as for Eq (9) can be used to formulate the inverse wavelet transform, expressed by: Hence, s P j and s Q j jointly form the synthesis part of the decomposition for the lossless reconstruction of the input semi-regular mesh J M . Note that the computation of the s A j and s B j involves the inversion of a large matrix, which makes the forward transform more complex than the inverse transform.

Lifting-based wavelet transform
As explained earlier, the filter bank implementation of multiresolution analysis is quite complex in the sense that the computation of analysis filters involve the computationally intensive inversion of large subdivision matrices. In this context, the lifting-based wavelet implementation (Schröder & Sweldens, 1995) provides a low complexity construction of multiresolution methods. In lifting-based multiresolution analysis, each scaling function  j i of the jth level exists so that    (Schröder & Sweldens, 1995). The refinement relation for the scaling functions is then: where l is the set which defines all linear combination of scaling functions and , j il p forms the entries of a matrix similar to P j . A similar refinement relation as Eq (11) In the forward transform, the first step is to produce a lower-resolution mesh j M starting from a higher-resolution version is predicted based on its low-resolution neighborhood in j M . After the prediction, an update step is used to modify the low resolution mesh j M . The update step is carried out on a pair 12 {,} cc of low-resolution vertices joined by a parent edge (Schröder & Sweldens, 1995) using the update weights  12 {,} aa . In general, the prediction and update weights only depend on the connectivity with respect to the vertex to be predicted. However, specific multiresolution analyses for which the weights depend on the specific resolution level and the underlying geometry can be also constructed (more details are given in Section 2.3).
The inverse transform can be formulated by following the forward-transform steps in the reverse order, i.e.:

Spatially Adaptive Wavelet Transform (SAWT)
As mentioned earlier, lifting-based transforms generally employ fixed prediction weights, independent of the spatial position and geometry around the vertex to be predicted. A simple observation reveals that a better prediction can result from adapting the prediction to the underlying geometry of the mesh. This argument is explained with a simple example: {} ii c lie on two different planes. In this situation, a prediction function for p c involving all eight coarser vertices will not be optimal and a better prediction could result by using 1 c and 2 c only. This is logical since p c lies on the edge formed by the vertex pair  

, cc
and is geometrically more correlated to vertices   prediction can be achieved if the prediction process is adapted to the local mesh geometry. Efficient prediction results in smaller energy of wavelet coefficients and hence an improved compression performance of the mesh coding system. To reverse the prediction operation, the decoder needs to know the weights used by the encoder for the prediction of each vertex p c . Since additional rate (compared to classical Butterfly) needs to be spent for coding the prediction weights, the total compression efficiency in the geometry adaptive case is a compromise between the bitrate saved due to the efficient prediction and the extra bitrate needed for signaling the prediction weights. In the following, a finite set of prediction filters is proposed in the context of spatiallyadaptive wavelet transforms (SAWT) (Denis et al., 2010a). The idea is to use one filter out of this set which best suits the geometry around the vertex to be predicted and which results in the smallest prediction error. A careful application of such an adaptive approach will provide an average rate gain if the reduction in the bitrate due to better prediction dominates the extra bitrate needed to signal the filter type to the decoder.
In a first step, the input semi-regular mesh is segmented into regions as follows. Let   For each region k , the wavelet analysis is performed by selecting one of the six candidates filters given below: for which different prediction filters will be selected.
Similar to (Chang & Girod, 2006), a filter candidate for a particular region k in J M , is chosen in an optimal distortion-rate (D-R) manner. More specifically, a predictor for each region k is selected such that the following Lagrangian cost function is minimized: where   kl R f denotes the rate necessary for encoding the filter index l used for prediction in the region k.

Scalable quantization of wavelet coefficients
In scalable mesh compression, the wavelet coefficients in the subbands are quantized using a generic family of embedded deadzone scalar quantizers (EDSQ) (Taubman & Marcelin, 2001), in which every wavelet coefficient X is quantized to: where    n denotes the quantization level.  n and  n denote the deadzone control parameter and the step size for any  0 n , respectively, with   0 2 n n and   0 2 n n , where  0 and  0 are the parameters for the highest rate quantizer corresponds to the well-known SAQ (Shapiro, 1993) in which the deadzone size is twice the step size  n for any n .

Wavelet coefficient histogram
In general, the observed histogram j wavelet subband is symmetric around its center of mass which is often zero or very close to zero. Moreover, the histogram is peaky around the mean and the frequency of occurrence decays as the magnitude of the coefficient's component increases. Fig. 5  In the literature, the observed histogram of any component of a wavelet subband is generally modeled using a zero mean generalized Gaussian (GG) distribution (Mallat, 1989), expressed by: where  ,     0,2 , is the shape control parameter.   0 is the scaling factor and , where  is the Gamma function. Note that, for   1 , Eq (17) transforms into a zero-mean Laplacian probability density function (PDF) given by: and for   2 Eq (17) corresponds to a zero-mean Gaussian PDF.
Although GG distributions closely approximate the observed histogram of wavelet coefficients, only approximate rate and distortion expressions for a uniformly quantized GG random variable are known (Fraysse et al,. 2008). The extension of these expressions to embedded quantization is not evident as the rate and distortion functions for such distributions are not easily tractable and can only be computed numerically. Moreover, computing these quantities gets very cumbersome due to the slow numerical integration of expressions involving a GG probability function, especially for   1 .

Proposed Laplacian mixture model
In order to avoid the aforementioned drawbacks of GG distributions, we propose a simple Laplacian mixture (LM) model which not only gives an easy closed-form derivation of the distortion and rate quantities but also better approximates the observed histogram of wavelet coefficients in the majority of cases. The proposed LM is a linear combination of two Laplacian PDFs, i.e., Note that The LM model is fitted over the observed data using the expectation maximization (EM) algorithm (Dempster et al., 1977) in order to determine the parameters  1 ,  2 and  . The E-step in the EM process calculates two responsibility factors 2 E and   0.9 , where  2 E is the estimated data variance. Histogram fitting for GG distributions is done using the bruteforce method where parameters 1  , 2  and  are exhaustively computed for a minimum KL distance.

Distortion-Rate (D-R) function
Closed-form expressions for the output distortion L D and the output rate L R of a Laplacian source quantized using an n level EDSQ are derived in the Appendix. In this section, we derive the D-R function for our proposed LM model. Since the distortion is a linear function of the source PDF, the output distortion LM D of the LM PDF for any quantization level n can be written as: corresponds to the deadzone cell). Since the LM model is symmetric around its mean,   kk PP . Note that the probability mass function (PMF) can be computed exactly due to the possibility of analytical integration of () LM f x . For the GG distribution, however, only numerical integration is possible.

Model validation
This section demonstrates that the proposed LM model is able to approximate the observed histogram and the observed D-R function of 3D wavelet coefficients more accurately compared to the commonly utilized GG distributions. For comparison purpose, results for the single Laplacian    5 illustrates that the proposed mixture model provides a better fitting probability function for the observed histogram compared to the Laplacian and GG distributions. This is especially true for the middle range positive and negative coefficients values -see Fig. 5. For the Rabbit mesh, LM gives only slightly better fitting than the other two models. However, for Dino, the LM can clearly model the fast decay of the observed histogram more accurately than the GG. The Laplacian PDF in this case only gives a very coarse approximation of the observed histogram.  (19)) for the normal (NOR) and the two tangential components (TAN1, TAN2) averaged over the five subbands. U-BF (Unlifted Butterfly), L-BF (Lifted Butterfly). In Table 1, the average KL divergence results for the Laplacian, GG and LM models for two non-normal (Venus, Rabbit) and three normal (Dino, Skull, Skredriver) meshes are shown. Each of the three coordinate components is considered separately. For each trial of Table 1, average is taken over five highest resolution subbands. For the large majority of cases, the LM model gives better fitting of the observed histogram than the competing GG model. Note that the Laplacian model gives always the worst fitting results. Also, the LM model gives equally good fitting for both normal (Nor) and tangential (Tan 1 and Tan 2) components. Superior histogram fitting results of our proposed model are also observed for the SAWT of Section 2.3. These results are not reported here due to lack of space.
In  Table 2. KL (%ME) for three resolution subbands averaged over the three coordinate components.
From Table 1, it is evident that on average the proposed LM model performs better than the GG and Laplacian models also in the ME sense. Better ME results are also obtained for SAWT (not reported here). Hence, the proposed LM model along with the derived D-R function is a better choice for modeling both the histogram and the D-R curve of mesh wavelet coefficients compared to the contemporary models. One notices that, a best histogram fitting in KL sense may not always yield the lowest ME . Table 2 reports the model validation results for different resolution subbands. For each trial the average is taken across the three spatial coordinate components. It is observed that the GG model performs slightly better for the low-resolution subbands of some meshes. The observed histograms in such cases are more Gaussian-alike, i.e., they have a round top. In general, the LM model faces difficulty in approximating such a round-top histogram due to the peaky nature of each of its Laplacian components; the GG fits well such histograms, as it corresponds to a Gaussian distribution for   2 . Nevertheless, the results show that, on average, the LM model outperforms the Laplacian and the GG models in KL as well as in ME sense.

Optimal embedded quantization
In this section, conclusions regarding the optimal EDSQ to be used in scalable wavelet-based coding of meshes are drawn. Let z denote the ratio between the deadzone size for n = 0 (see Eq. (16)) and the step size for  0 n of a general EDSQ. The total average signal-to-noise ratio (SNR) difference which is utilized to measure the performance gap of different embedded quantizers is defined as: which is computed over a rate range  for N rate points, where () SNR R denotes the discrete SNR-rate function. The   2 10 10 log ( ) SNR D is computed in dBs, where D is the total distortion in the transform domain. The difference in SNR is computed relative to the uniform embedded quantizer (UEQ), i.e.,  1 z . SNR for five embedded deadzone quantizers is plotted in Fig. 7. over a wide range of standard deviation ratios   21 . In Fig.  7., the commonly observed proportion   0.9 is considered, as mentioned in Section 3.2.
We determined experimentally that at lower standard deviation ratios, SNR is positive and the UEQ is optimal for  21 120 . . In general, small standard deviation ratios correspond to  close to 1 , observed in nonnormal meshes, while higher ratios correspond to   1 , observed in normal meshes. These results show that one cannot determine a single embedded quantizer that provides the best performance for all 3D meshes. However, an optimal quantizer per wavelet coordinate can be determined based on the corresponding   21 extracted from the model.
Overall, for  21 120 , the difference between SAQ and the UEQ is significant, and hence UEQ is the optimal choice. For  21 120 , SAQ is not always the optimum, but lies not far from the optimum. Given the fact that SAQ is closely linked to bit-plane coding and that it can be implemented using simple binary arithmetic, one concludes that SAQ is not an optimal, but an acceptable solution in scalable coding of meshes.

Analysis of wavelet coefficient dependencies
Similar to images, parent-children and neighboring wavelet coefficient dependencies exist in wavelet decomposed mesh structure. In Fig. 8 (middle, right), the positions of the wavelet coefficients at different levels of the transform are shown with the help of white and dark circles. In particular, wavelet coefficients have a one-to-one correspondence with the edges of the coarser mesh. For each wavelet coefficient there are rings of neighboring coefficients which lie in the same wavelet subband -see Fig. 8 (right). Also, a set of four wavelet coefficients have a parent coefficient at the next coarser resolution -see Fig. 8 (middle, right). Fig. 8. Parent-children and neighboring wavelet coefficients: actual mesh (left); coarser meshes after one (middle), and after two wavelet decomposition levels (right).
Statistical intraband dependencies exist between neighboring coefficients of each resolution level. The main reason for the existence of these dependencies is the smoothness of the surface. Wavelet coding paradigms that exploit the intraband dependencies between the wavelet coefficients are known as intraband wavelet codecs such as block-based coding techniques (Munteanu et al., 1999a), quadtree coding approaches (Munteanu et al., 1999b), and the EBCOT codec employed in the JPEG-2000 scalable image coding standard (Taubman, 2000).
Statistical dependencies also exist between the parent and descendants (children) due to the natural decay of the coefficients' magnitude for increasing frequencies. In other words, if a parent coefficient magnitude is below a certain threshold, then there is a high probability that the magnitude of its descendants will be also below this threshold. This corresponds to the so-called zerotree-model, firstly introduced by Shapiro in (Shapiro, 1993). The wavelet coding paradigms that exploit the parent-children dependencies are known as interband wavelet codecs.
Finally, there is a third category of coding paradigms, exploiting both the interband and intraband statistical dependencies between the wavelet coefficients. They are generally known as composite codecs, EZBC (Hsiang & Woods, 2000) and the ECECOW approach of (Wu, 1997) are typical examples of codecs in this category.
In the following, an information theoretical analysis of the aforementioned coefficient dependencies is presented. Our aim is to single out the type of dependencies which can ensure best compression performance in the context of wavelet-based mesh compression.

Mutual information analysis
The mutual information is the reduction in the entropy of one random variable due to the knowledge of the other random variable.
It is known that  (,) (,) IXY IYX . In the wavelet domain, we define the following mutual information quantities: (; ) X IXP : denotes the mutual information between a wavelet coefficient X and its parent coefficient X P .
: denotes the mutual information between a wavelet coefficient X and its neighboring wavelet coefficients : denotes the composite mutual information.
From the basics of information theory (Cover & Thomas, 1991), we know that: , we need to estimate the joint PDF (, ) x pxn which can have high dimensionality depending on the number of considered neighbors. Since the amount of data needed to accurately estimate a PDF increases exponentially with its dimensionality, it is difficult to reliably estimate a high-dimensional PDF. To alleviate this problem, the reduction in dimensionality as proposed in (Liu & Moulin, 2000) is used here. We summarize the neighborhood of X through a so-called summarizing function  n () X Tg . This function maps the neighboring wavelet coefficients to a single value. We note that such a many-to-one summarizing function cannot increase the mutual information, i.e., (; ) Equality in the above equation holds if  n X TX forms a Markov chain. The summarizing function used in our analysis is: Due to this summarizing function, it is sufficient to compute the joint PDF (,) p xt , t is a realization of the random variable T , instead of (, ) x pxn , for the estimation of the intraband mutual information n (; ) X IX .
In our analysis, the mutual information for the defined quantities is estimated using the adaptive partitioning method (Darbellay & Vajda, 1999) instead of the traditional histogram method. This is because the histogram method highly depends on the bin size and for a small bin size there may not be sufficient number of observations in some bins to make a correct estimate. The adaptive partitioning method (Darbellay & Vajda, 1999) on the other hand, ensures that there are always sufficient numbers of observations in each bin, and provides reliable estimates of the mutual information.  Table 3. Average mutual information in bits for several non-normal and normal meshes. Table 3 shows the average mutual information results for interband, intraband and composite dependencies for various mesh models. Since in mesh coding three different components need to be coded for each vertex position in space, the average mutual information

Butterfly
is reported instead of the mutual information for the three components individually. It is observed from Table 3 that for both normal and non-normal meshes mutual information of interband models is the least, independent of the wavelet transform employed. On the other hand mutual information for intraband models is significantly higher than for the interband models. Finally, composite models which gather the characteristics of both interband and intraband models exhibit even higher mutual information than interband or intraband models alone. Mathematically we can summarize our numerical findings as: Experimental results for the mutual information based estimation of interband, intraband and composite dependencies seem to indicate that exploiting the composite dependencies should be preferred. Additionally, it is important to point out that favoring intraband over zerotree-based interband models brings along the additional benefit of resolution scalability. Specifically, by following an intraband codec design, only those wavelet subbands that are needed in order to reconstruct a target mesh resolution-level need to be encoded, while the others can be discarded. This does not hold in case of interband and composite codec designs, due to the tree-structures that span all the wavelet decomposition levels. Since composite models cannot be discarded altogether due to their highest mutual information property, a careful implementation of a composite mesh coding system needs to be carried out into order to get the benefit of both the higher compression efficiency and the resolution scalable decoding at the same time.
Finally, it is important to point out that the differences in terms of mutual information do not give any indication about the final performance differences between interband, intraband and composite coding systems. Hence, an actual development and comparison of such coding systems is needed in order to experimentally validate the conclusions of this information-theoretic analysis of wavelet-based mesh coding designs, which is presented next.

Scalable mesh compression overview
In this section, we give a brief overview of the scalable mesh compression systems. Based on the design choices established earlier, we designed intraband and composite mesh coding systems which provide state-of-the-art compression performance, together with resolution as well as quality scalability of the compressed mesh.

Progressive Geometry Compression (PGC)
The first scalable wavelet-based geometry compression technique is the progressive geometry compression (PGC) codec proposed by Khodakovsky et al. in (Khodakovsky et al., 2000). PGC makes use of the well-know zero-tree coding (Shapiro, 1993) of wavelet coefficient's bitplanes in order to encode the decomposed mesh structure. Significant improvements in the compression performance against the contemporary scalable as well as non-scalable mesh coding systems were reported in (Khodakovsky et al., 2000). However, a major drawback of PGC schemes is their inability to provide resolution scalability. This is caused by the zero-tree structure which, for a given bitplane, spans all the wavelet decomposition levels. For a detailed understanding of the PGC system we refer to (Khodakovsky et al., 2000).

Scalable Intraband Mesh Compresion (SIM)
Despite of the great success of zerotree-based coding techniques in image coding, the choice of an interband codec design is not necessarily the best option in the context of scalable mesh coding. This was illustrated in Section 4 where different types of dependencies among wavelet coefficients were studied. Based on this analysis, we opt for an intraband dependency model in our codec design. As mentioned before, favoring intraband models over interband models brings along the additional benefit of resolution scalability. Specifically, by following an intraband codec design, only those wavelet subbands that are needed in order to reconstruct a target mesh resolution-level need to be encoded, while the others can be discarded.
In the designed scalable intraband mesh (SIM) compression system (Denis et al., 2010b) each resolution subband is encoded independently of the others. Similar to (Shapiro, 1993), SAQ is applied to each resolution subband to determine the significance of the wavelet coefficients with respect to a series of monotonically decreasing thresholds. Based on the significance outcome, a tree node is split into eight equal volume nodes. The resulting octree nodes may contain an unequal number of wavelet coefficients. In general, the number of coefficients in all nodes of a same tree-depth is roughly the same. This way, an octree is constructed for each resolution subband, wherein the depth of the tree (number of levels in the octree) is equal to the number of bitplanes of the subband. All magnitude bitplanes are sequentially coded using the non-significance, the significance and the refinement coding passes. For a detailed presentation of the SIM codec the interested reader is referred to (Denis et al., 2010b).
Using the octree-based bitplane coding, separate symbol streams are first generated for all bitplanes of each resolution subband. Depending on the type of scalability, i.e., resolution or quality scalability, the encoded symbol streams are entropy coded using a predefined progression order of bitplanes. For quality scalability, bitplanes of certain significance, from all resolution subbands, are first encoded before encoding the bitplanes of lower significance. However, in resolution scalability mode, all bitplanes of a lower resolution subband are progressively encoded before encoding the next higher resolution subband.
We compared the SIM codec with the PGC codec for both normal and non-normal 3D meshes. The decoded meshes are compared against the original semi-regular input meshes using the peak signal-to-noise ratio (PSNR) as the distortion metric, which is defined as: where peak and RMS denote the size of the bounding box and the root mean squared error calculated on the distances between the decoded vertex positions with respect to the original ones, respectively. The averaged gain in PSNR when compressing the Venus and Bunny meshes goes up to 2.22 dB and 2.35 dB, respectively. One may also notice the increasing performance difference with increasing bitrates; this indicates that the SIM coder tends to code the high frequency information more efficiently. For the spatially adaptive wavelet transform (SAWT) the compression results are reported in (Denis et al., 2010a).  10 shows compression performance plots for two normal meshes, Skull and Dino. One notices that at low bitrates, PGC tends to compress better. However, the ability of SIM to capture and code more efficiently the high-frequency components is noticeable at high bitrates and leads to an improved performance when compared to PGC.

Composite Context-conditioned Compression (3xC)
The mutual information analysis presented earlier showed that the composite dependencies between the wavelet coefficients are by far the strongest. However, one may notice that, employing composite models may hinder, similar to interband models, the possibility of providing resolution scalability. Thus one must be careful in exploiting the parent-children dependencies within composite models. A careful observation reveals that exploiting parent-children dependencies in a causal fashion (Denis et al., 2010b) does not limit resolution scalable decoding of the compressed mesh. Following this observation, we proposed a scalable composite mesh compression system in (Denis et al., 2009), (Denis et al., 2010b). The bitplane coding modules of the SIM codec and the 3xC codec are identical. The two designs differ at the entropy coding level. In particular, for 3xC, parent coefficient based context-conditioning is employed in the entropy coding module. For context-conditioning, significant, non-significant as well as sign information is entropy coded using the designed context tables. The refinement information is encoded without context-conditioning; this is because including the parental information when entropy coding the refinement symbols does not improve compression performance. For a detailed presentation of the 3xC codec the interested reader is referred to (Denis et al., 2009). Fig. 9 also depicts the PSNR curves computed for the non-normal Venus and Bunny meshes using our implementation of the un-lifted butterfly based 3xC mesh compression system. The figure clearly demonstrates that, when dealing with non-normal meshes, 3xC systematically yields superior performance compared to PGC as well as SIM. In the case of normal meshes (Fig. 10) our coder employs the same transform as PGC. Both codecs perform the same at very low bitrates. However, overall, 3xC yields the best compression performance. 3xC gives approximately equivalent results when compared with the intraband SIM codec for normal meshes. This is because the context-conditioning is only possible for the normal component of vector valued wavelet coefficients. Overall, it is clear that the proposed 3xC codec produces similar, and in almost all cases, superior performance compared to PGC and SIM codecs.

Visual comparison: PGC vs 3xC
Visual comparisons of Bunny and Skull meshes, compressed and reconstructed using 3xC at different bits per vertex (bpv), are presented in Fig. 11 and Fig. 12, respectively. The colored regions highlight the distortions introduced by lossy compression. For low-to-medium bitrates, the pure red color indicates areas where the distance between the original and decoded vertex is larger than 0.1% of the diagonal of the bounding box of the semi-regular mesh. For high bitrates, the distortion is visualized with respect to 0.02% of the diagonal. The mesh is shaded greener as the distortion lowers, with pure green indicating no distortion.
When visually comparing the compressed Bunny and Skull meshes produced by 3xC and PGC, it is very clear that 3xC yields superior performance for all bitrates. Taking the result at 0.050 bpv as an example, we observe that many areas which are shaded red for PGC are green for 3xC. At high rates, the differences between the mesh geometries may not be visually significant, yet the colors reveal that 3xC is able to approximate the original mesh much more accurately compared to the PGC system.  The visual comparisons of the normal mesh Skull at different bpv are shown in Fig. 12. Though, at first glance it may appear that both codecs perform very similar, small differences are noticeable when investigating the meshes more closely. When examining the comparison at 0.036 bpv, we notice that the PGC codec preserves more details in Skull's teeth. The green shade for 3xC at rate 0.179 bpv, however, seems more pure compared to PGC for which it is rather yellowish green. We also observe that no red regions are present for 3xC at rate 1.073 bpv, whereas some are visible for PGC at the same rate.

Conclusions
In this book chapter, we propose a constructive methodology for the design of scalable wavelet-based mesh compression systems. Our design strategy differs from conventional designs which simply opt for reusing methods from wavelet-based image coding for the design of mesh coding systems. In particular, our methods are motivated by an informationtheoretic analysis of the statistical dependencies between wavelet coefficients which shows that, intraband dependencies are systematically stronger than interband ones for both normal and non-normal meshes, and that composite models are the best. We also investigate the optimality of successive approximation quantization, commonly used in scalable compression, in the context of wavelet-based mesh compression. Using a Laplacian mixture model, it is shown that successive approximation quantization is an acceptable, but in general not an optimal solution. Anchored in these results, novel intraband and composite coding systems are presented which improve the state-of-the-art in scalable mesh compression, both in terms of scalability and compression efficiency.

Appendix
The output distortion L D of a Laplacian PDF, quantized using an n level EDSQ and reconstructed using midpoint reconstruction, can be written as: Similarly, the output rate L R of a Laplacian PDF, quantized using an n level EDSQ can be written as: ( The use of the wavelet transform to analyze the behaviour of the complex systems from various fields started to be widely recognized and applied successfully during the last few decades. In this book some advances in wavelet theory and their applications in engineering, physics and technology are presented. The applications were carefully selected and grouped in five main sections -Signal Processing, Electrical Systems, Fault Diagnosis and Monitoring, Image Processing and Applications in Engineering. One of the key features of this book is that the wavelet concepts have been described from a point of view that is familiar to researchers from various branches of science and engineering. The content of the book is accessible to a large number of readers.