Structural and Temporal Analysis for Time-Varying-Meshes

Abstract

This thesis proposes algorithms for structural and temporal analysis for Time- Varying-Meshes (TVMs), through hierarchical skeleton-based mesh segmentation and surface curvature matching. The motivation is to understand and retrieve structural information while establishing spatio-temporal correspondence in TVMs. With that, we can also achieve markerless motion transfer from humans onto synthetic models. The thesis also introduces a parts-based skeleton extraction method using geodesic distances to handle badly-defined meshes that are common in TVMs. Thus the algorithm enables tracking of highly deformable models with arbitrary genus. The proposed framework is a recursive system that iterates between hierarchical segmentation on minimum distance satisfaction and skeleton realignment to refine a kinetic model for each TVM frame. Results show that the segmentation is stable and successful motion tracking is achieved from TVMs. We also explore the use of surface curvatures in Iterated Closest Point (ICP) matching for more refined motion understanding on the segmented results. Results show that processing time can be reduced by 80% while retaining the same level of accuracy. Multi-temporal registration, where the data in the previous and the following frames are used in the 3D shape registration, is used to increase robustness against potential bad registration as ICP is sensitive to outliers and deformation. It achieved an angle disparity reduction of average 29% with two additional time registrations.