서강대학교

초록/요약

We present a new local shape blending system that maps a sparse configuration of facial markers captured from an actor on to target meshes, by blending highly detailed target key meshes. The traditional local blending computes weight vectors of each region on the mesh with respect to the key meshes. Here all the vertices in a region share the weight vector of the region. While respecting the coherence or quazi-rigidness of control points within each region, this method decouples the natural correlation between regions. This problem has been partly addressed by a recent soft-region method which computes the weight vectors of each control point independently of each other, and then computes the weight vector of each vertex by scattered data interpolation over the weight vectors of the control points. But this method ignores the coherence that may exist among control points. The underlying issue here is how to compare the observed con guration of control points and their con guration computed by blending key meshes. We solve the problem by comparing the observed configuration and the computed con guration region by region, by designing regions to be coherent weighted regions. These regions are de ned by weight functions that use the generalized distance between control points called the coherency-based distance: two control points are more coherent or more closer when they are closer to each other in geometric distance, and they have more similar blending weight vectors. To use coherent weighted regions systematically, we formulate local shape blending as a problem of nding an optimal con guration of blending weight vectors, assuming that the con gurations of weight vectors obey a Markov Random Field.