Parametric Representation Of Full Human Dynamical Geometries Based On Sparse Localized Decomposition

Real-time acquisition of 3D dynamic geometries has achieved rapid progress recently,which makes dynamic geometry processing progressively becoming the important contents of computer graphics and computer animation.Editable and compact representation is a core issue in dynamic geometry processing and it lays a foundation for effective storage,intuitive reuse and fast reconstruction of 3D dynamic geometries.This paper investigates parameterized representations of two kind of dynamic geometries: the first one is the animated mesh sequence of a single object,and the other is animated meshes of multiple people as a whole.Sparse localized bases are capable of capturing deforming regions and motion modes in a mesh sequence,however,existing approaches may fail to extract sparse localized components effectively supporting semantic editing when there exists global rotation or large-scale local rotation in the sequence.To address the issue,this paper proposed an articulated-motion-aware sparse localized decomposition approach.Given an animated mesh sequence,it first collects the variation of edge length and dihedral angle of each pose frame with respect to the reference pose into a vector,and then puts vectors of poses together to form an edge length and dihedral angle variation matrix;it then extracts a set of sparse localized components which can effectively capture the variation of edge lengths and dihedral angles of different poses;shape editing of these poses can then be achieved by changing the blending weights of these components.Theoretically,edge length and dihedral angle are invariant up to rigid motions,therefore they remain unchanged when the object plays a global rotation;moreover,the dihedral angle of arbitrary edge is always limited in [0,360°] during motion and hence large-scale deformations will not lead to ambiguity of its value.Experimental analysis also demonstrates that the proposed representation effectively tackles the issues of large-scale deformation and global rotation,and simultaneously is able to support both articulated and non-rigid motions(cloth wrinkles,muscle deformations,and facial animations)editing.Researchers have acquired a great number of 3D geometries of poses,facial expression and hand gestures due to the particularity of human.Compared to animated meshes of a single object,these geometries have one more dimension—shape variation.To parametrizing this data as a whole,existing methods are prone to result in shape distortion near joints due to taking advantage of the skeleton skinning to drive motions.Other the other hand,it is still a challenge to express body postures,facial expressions and hand gestures(we name it human complete whole dynamic geometries)in a unified framework.We propose a novel human complete whole dynamic geometry parameterization model——PANOMAN to attack the issue.Similar to existing methods,PANOMAN makes use of PCA bases to represent different shapes.However,it employs an improved articulated-motion-aware sparse localized decomposition to extract motion bases.Owing to the locality of motion bases,it is possible to construct training datasets via synthesizing existing facial expression,hand gesture and body pose databases instead of physically acquiring.This greatly reduces the burden of collecting training data.We also investigate the reconstruction framework of 3D human dynamic geometries and design a common-used reconstruction optimization model.We demonstrate via experiments that PANOMAN enable the effective representation of human dynamic geometries integrating different type and different scale human motions together;furthermore,its reconstruction results are more accurate than those of the state-of-the-art methods in generating large-scale twisting poses and unusual hand gestures.