Global Alignment for Dynamic 3D Morphable Model Construction

I. ABSTRACT We propose a novel 3D Dynamic Morphable Model which utilizes the Dynamic 3D FACS Dataset (D3DFACS)[1]. Compared to the technique proposed in [1], the model provides a more accurate global representation and improved methods for aligning independent sequences of facial expressions. Firstly, we find the rigid correspondence between raw meshes by using Coherent Point Drift (CPD)[2]. In order to obtain more precise correspondences on UV texture maps, we then use a novel hybrid registration approach combining Thin-plate Splines (TPS) and optical flow [3]. We compare this approach of alignment of independent dynamic sequences against standard optical flow methods proposed in [1]. After both rigid and nonrigid alignment, a 3D dynamic morphable model is constructed using Principle Component Analysis (PCA). We aim to use the model to assess the ability of 3D dynamic morphable models on speech replications for talking head and other animation applications.