Data-driven weight optimization for real-time mesh deformation

Abstract

3D model deformation has been an active research topic in geometric processing. Due to its efficiency, linear blend skinning (LBS) and its follow-up methods are widely used in practical applications as an efficient method for deforming vector images, geometric models and animated characters. LBS needs to determine the control handles and specify their influence weights, which requires expertise and is time-consuming. Further studies have proposed a method for efficiently calculating bounded biharmonic weights of given control handles which reduces user effort and produces smooth deformation results. The algorithm defines a high-order shape-aware smoothness function which tends to produce smooth deformation results, but fails to generate locally rigid deformations.To address this, we propose a novel data-driven approach to producing improved weights for handles that makes full use of available 3D model data by optimizing an energy consisting of data-driven, rigidity and sparsity terms, while maintaining its advantage of allowing handles of various forms. We further devise an efficient iterative optimization scheme. Through contrast experiments, it clearly shows that linear blend skinning based on our optimized weights better reflects the deformation characteristics of the model, leading to more accurate deformation results, outperforming existing methods. The method also retains real-time performance even with a large number of deformation examples. Our ablation experiments also show that each energy term is essential.