Efficient Pooling Operator for 3D Morphable Models.

Abstract

Learning the latent representation of three-dimensional (3D) morphable geometry is useful for several tasks, such as 3D face tracking, human motion analysis, and character generation and animation. For unstructured surface meshes, previous state-of-the-art methods focus on designing convolution operators and share the same pooling and unpooling operations to encode neighborhood information. Previous models use a mesh pooling operation based on edge contraction, which is based on the Euclidean distance of vertices rather than the actual topology. In this study, we investigated whether such a pooling operation can be improved, introducing an improved pooling layer that combines the vertex normals and adjacent faces area. Furthermore, to prevent template overfitting, we increased the receptive field and improved low-resolution projection in the unpooling stage. This increase did not affect processing efficiency because the operation was implemented once on the mesh. We performed experiments to evaluate the proposed method, whose results indicated that the proposed operations outperformed Neural3DMM with 14% lower reconstruction errors and outperformed CoMA by 15% by modifying the pooling and unpooling matrices.