Matching and retrieval of 3D deformable models

Abstract

Due to the popularity of computer graphics, 3D games and CG movies, a large number of 3D models have been created. Many of them are freely available on the Internet and can be found in many personal homepages, repositories and even libraries. Though the number of 3D models keeps on increasing, 3D models are time and effort consuming to build. The group of deformable models (or so called animating models or articulated models) is especially a good example. These models are usually high quality meshes that require weeks of work to construct. As such, it would be advantageous to build models by reusing or adapting from existing ones. To facilitate reuse and sharing, research on 3D model matching and retrieval becomes essential. Existing content-based 3D model matching methods can be categorized into four major approaches: geometry-based, transform-based, image-based and topology-based. Among these four approaches, only topology-based matching methods can handle highly deformable models, i.e., models representing the same object but in different postures. In this research project, we target ourselves at the topology-based retrieval techniques. We believe that a successful topological retrieval system will facilitate sharing and reuse of these models, and it will greatly benefit the game and movie production industry in particular by reducing the overall production cost. Topology-based methods analyze 3D models based on their skeletal information. Since skeletal features are invariant to models’ postures, topology-based methods can thus handle deformable models and the matching results agree with human intuition. However, there are only several topology-based retrieval methods, and most of them require high computational cost. Researching for a fast and accurate method to retrieve deformable models becomes our ultimate goal. In this thesis, we explore these methods and also take a novel approach to analyze the same problem. Instead of using explicit skeleton or multi-resolution reeb graph (MRG), we propose the use of topological characteristics to represent a model. Since our method makes use of the flow and transportation algorithm (Earth Mover Distance) for model matching, it is shown to be fast and efficient. As a comparison, our proposed method is 4.5 and 15 times faster than the method MRG in feature extraction and matching, respectively. Though our method does not compare models by skeletons, the additional high-dimension geometric features that we introduce are capable of discriminating different models of similar and dissimilar skeletons. From our experiments, our method also outperforms MRG in accuracy and can even achieve a precision of 0.77 at a recall rate of 1.0.

Date of Award	15 Feb 2005
Original language	English
Awarding Institution	City University of Hong Kong
Supervisor	Chong Wah NGO (Supervisor) & Rynson W H LAU (Co-supervisor)