CODDYAC: CONNECTIVITY DRIVEN DYNAMIC MESH COMPRESSION Libor Váša, Václav Skala {lvasa|skala@kiv.zcu.cz} Department of Computer Science and Engineering Faculty of Applied Sciences University of West Bohemia ABSTRACT Compression of 3D mesh animations is a topic that has received increased attention in recent years, due to increasing capabilities of modern processing and displaying hardware. In this paper we present an improved approach based on known techniques, such as principal component analysis (PCA) and EdgeBreaker, which allows efficient encoding of highly detailed dynamic meshes, exploiting both spatial and temporal coherence. We present the results of our method compared with similar approaches described in literature, showing that using our approach we can achieve better performance in terms of rate/distortion ratio. Index Terms — Dynamic mesh, compression, PCA, EdgeBreaker, coherency exploitation, entropy 1. INTRODUCTION Dynamic mesh is a common term used for a series of static triangular meshes that represents a development of some surface in time. Usually, two additional assumptions are made about the dynamic mesh: • every mesh in the series has the same connectivity, i.e. there is an one-to-one correspondence of vertices from frame to frame of the animation • the animation represents some physical process, i.e. there are no sudden changes in the geometry of the subsequent frames of the animation. The problem of compressing such data structure has been addressed in the past by various approaches. One class of algorithms is based on various spatio-temporal predictors, which are an extension of the spatial predictors, such as the parallelogram predictor [1]. The spatio-temporal predictors apart from using positions of vertices in the current frame also use vertex positions form one or more preceding frames. One of the first attempts in this field are the two spatio temporal predictors introduced with the DynaPack algorithm by Ibarria and Rossignac [2], the ELP predictor, and the Replica predictor. The connectivity driven predictor proposed by Stefanoski [3] is another example of a spatio-temporal predictor. The approach suggested by Mueller et al. [4] can be seen as an augmentation of the predictor based technique by using spatial subdivision by an octree structure according to the size of prediction residuals. A slightly different approach has been proposed by Payan et al. [5]. Their approach is targeted on exploiting the temporal coherence by wavelet decomposition of the vertex trajectories, followed by a rigorous bit-allocation process that optimizes the rate-distortion ratio. Alexa and Mueller [6] have suggested a PCA-based approach based on reducing the space of shapes (frames). Their approach first requires a rigid motion compensation (the least-squares method is used to find an affine transform that best fits the given frame with respect to the first frame of the animation), and each frame is then expressed as a linear combination of uncorrelated basis of the space of possible shapes, called EigenShapes. The method is based on singular value decomposition (SVD), as the space dimension (number of vertices) is much greater than the number of samples (number of frames). Several improvements of the original Alexa’s method have been proposed. First, Karni and Gotsman [7] have suggested using linear prediction coding (LPC) on the PCA coefficients in order to exploit temporal coherence of the data. Subsequently, Sattler et al. [8] have proposed an application of PCA onto trajectories instead of shapes, and combining the approach with spatial clustering which allows exploitation of spatial coherence. A combination of PCA and local coordinate frame (LCF) has been suggested by Amjoun [9]. The approach is based on spatial clustering driven by the magnitude of vertex coordinates expressed in each cluster’s LCF. Recently, Mamou et al. [10] have proposed a method based on automated skinning, which also uses the spatial clustering. An affine transform that best fits each cluster movement in time is found, and each vertex is described by a weighted sum of the transforms associated with the cluster the vertex belongs into, and of the transforms assigned with the neighboring clusters. More detailed description of the compression approaches can be found in [11]. Generally, the methods can be classified according to what approach is used to 3DTV Conference proceedings, IEEE ISBN 978-1-4244-0722-4, 2007