Volume-based 3D Shape Retrieval And Morphing

The number of 3D models is increasing rapidly along with the development of multimedia, animation and CAD technologies. Content-based 3D shape retrieval is a research focus because 3D models lack text information and there are usually complex. Aided by retrieval technologies, case-based shape innovation is a common way in some applications, e.g., product design and amination invention. To improve the applicability, it is necessary to unifiy the shape representation during retrieval and morphing computation.A binary voxelization method, i.e., error-bounded soild voxelization for polygonal models based on heuristic seed filling, is proposed to aid volume-based retrieval. The method is a balance among robust, efficient and error. It sovles different problems of other solid voxelization methods, such problems including low efficiency, impratical requirement to polygonal models and lacking means to control error. The method consists of three stages: Primary Binary Volume Model (PBVM), Boundary Binary Volume Model (BBVM), Revised Binary Volume Model (RBVM). RBVM discovers heuristic knowledge from PBVM and BBVM. Heuristic seed filling improves efficiency remarkably and keeps the robustness of conventional seed filling.To improve the applicability of retrieval, volume-ratio is used to extract feature vector. Continue principal components analysis is applied to polygonal models to normalize postion before feature vector computation. Volume ratio is computed by bin-subdividing the bounding box of the polygonal model along axes recursively. Based on such volume ratios, similarity is obtained by weighted sum. Two feedback mechanisms are proposed to improve retrieval result, including single weight feedback and first order statisfication evaluation. Based on feature vector computation, efficient feature indexing is propsed as well and pre-filtering is used to improve retrieval speed.To do distance-based morphing, distance transformation is proposed based on fast marching method. At first, supercover voxelization is appled to continue surface to install reference zone. The distance field is propagated from reference zone by 6-conntected narrow band and heap sorting is used to select minimal distance from all candidate voxels in the propagating process. The reference of each voxel is recorded to setup the complete distance field representation which is the basis for computation of higher resosultion distance fields.Level-set method is adopted to do morphing since its ability to deal with shapes with different topological properties. Single-sided active set is proposed to improve the efficency of sparse field algorithm, which is a fast numerical solution of level-set. Other layer sets are defined by topology accordingly. Such definitaions not only decrease the number of active voxels but give a more efficient way to update other layer set voxels. Two smoothing methods, i.e. averaging&translation and narrow band evolving&back, is proposed to complement the alias arised by approximate computation of Euclidean distance in the evolving process.