3D object representation and analysis with application in object simulation and retrieval

三維模型的表達, 分析及其在三維仿真與模型檢索中的應用

Student thesis: Doctoral Thesis

View graph of relations


  • Qizhen HE


Awarding Institution
Award date4 Oct 2010


3D applications have been widely used in many fields ranging from industrial CAD/CAM design to medical simulation of organ deformation and to the entertainment industry in our daily life. More and more applications based on 3D scene come cross and give customers much better experiences. Although people have already got some sophisticated methods to reconstruct 3D objects from a set of 2D images, how to represent and analyze the objects in more effective ways still remains a problem, which is now becoming a hot research topic related to the areas of computer vision and machine learning, to which many researchers devote themselves. By taking advantage of the statistical and machine learning methods, the thesis provides effective solutions to applications of two categories: learning deforming patterns to describe the object‘s physical behavior for object simulation to and extracting descriptive features to describe the object‘s geometrical characters for objects retrieval. For object simulation, previous work focused on adapting physical analysis such as mass spring model and finite element analysis to maintain the object‘s material property. Such approaches typically require high computational costs. In this thesis, a novel Statistical Deformable Model (Mr-SDM) is proposed to learn the deforming patterns related to the material properties from a set of man-made training samples generated through Finite Element Method (FEM). By combining the approaches of SDM and FEM, Mr-SDM speeds up the process of physical analysis. The distinctive advantage of our approach is that Mr-SDM moves the high computational cost of FEM to become offline training while at the same time, solves the small sample size problem of traditional SDM. The advantage of Mr-SDM is verified and evaluated by applying it to simulate the deformations of generic objects with different shapes and material parameters, as well as to predict the facial changes in the orthognathic surgery. For object retrieval, currently the objects are represented by a set of raw features such as spin images and relative angle context distributions which measure the statistical properties of the objects and lead to tremendous data redundancy. In this thesis, a novel high level representation of 3D objects called visual topics are proposed to extract the shape characters from the raw features by clustering the nearby vertices or views with similar behaviors. To understand how such visual topics deform among different resolutions, and more importantly, the topics from which resolution are the most descriptive and leads to the best retrieval results, we further extended visual topics to multi-resolution to describe the object at different resolutions by smoothing the object gradually into a sphere. We applied the topic representation for 3D object retrieval and evaluated the performances with different similarity measure. The experiment results demonstrate the visual topic representation of the 3D objects is not only effective (for its good performance in object retrieval) and economic (for its reduction in raw features‘ redundancy) but also promising (for it is associated with visually semantic meanings over mesh representation of the object). Keywords: 3D Object, Deformation, Finite Element Method, Statistical Deformable Model, Feature Extraction, Shape Topics, View Topics, Multi-Resolution Feature

    Research areas

  • Three-dimensional imaging, Digital techniques, Image processing