| With the innovation of techniques for 3D modeling and the development of internet, it becomes easier and easier to obtain and create 3D models. The rapidly growing amount of 3d models makes content-based 3D model retrieval be a very important research issue.A traditional method of retrieval of 3D model is text-based search which searches a model in database via its textual annotation. Text-based search can be effective when textual annotation and consistency of annotation exist. Since textual annotations are usually generated manually, they are subject to a person who creates them. This fact causes inconsistency issues. It is not proper to describe many properties by text, such as shape and surface texture. Meanwhile, it is a fact that manual annotations are time consuming. In addition, it is not all of models contain corresponding textual annotation. In contrast, content-based 3D model retrieval methods, which use shape properties of the 3D models to search for similar models, work better than text-based ones. Thus content-based 3D model retrieval becomes a key research issue.Though there have been a great number of achievements of 3D model retrieval both in theoretical and practical fields, it is still necessary to find ways to improve current methods effectiveness and to explore new methods for 3D model retrieval. As there are more and more 3D models with color, texture or reflective value properties, 3D model retrieval based on surface properties becomes a valuable research issue. Works and contributions of the thesis are stated as follow:(1). It is important for content-based 3D model retrieval to increase effectiveness of methods which have been proposed. The thesis proposed a common feature based 3D model retrieval method with analysis for different feature vectors extracted from 3D model. There are many methods to describe feature vector from 3D model for content based 3D model retrieval. Considering some basic facts that feature vectors extracted from 3D models should be similar when these models are similar, it is can be drawn that there are common parts in different feature vectors. By analyzing the similarity for different feature vectors of models in the same class, common parts and distinct parts can be found. Then a signature of all models in this class, which is called class feature vector, is generated by combining common parts and distinct parts in some way. In the process of 3D model retrieval, several classes which are probably related with the query model could be retrieved firstly by comparing the feature vector of query model with the class feature vectors of classes in search space. Then, related models are retrieved from these classes. This method is more effective and efficient than traditional method. Furthermore, it is a universal method and can be used with many feature vectors.(2). Extending Generic Fourier Descriptor to 3D space, a z-axis rotation invariant 3D shape descriptor, SFT, is proposed. SFT is based on spherical sampling and 3D fourier transform. This descriptor is concise and robust to model noise and invariant to rotation around z-axis. This method can give good retrieval effectiveness without pose normalizing in general 3D model database and can give more precisely descriptor with invariance to rotation around z axis in peculiar retrieval process in which a stable axis of every model in database is given.(3). By studying surface properties based 3D model retrieval, three different surface properties based descriptors, Color Distribution, Spherical Color Distribution and Colored Spherical Harmonic, are proposed. Color Distribution descriptor is obtained from statistics of color values which are sampled from models in quantized color space. Spherical Color Distribution descriptor is obtained from statistics of color values sampled from the sphere that is generated of unbiased sampling from model. Colored Spherical Harmonic descriptor is obtained by using spherical harmonic on separated RGB spherical functions of the sampled sphere. These proposed descriptors have been evaluated on the colored model database. Experimental results show that these descriptors give good retrieval performance in surface properties based retrieval.(4). This thesis initially explores the hybrid descriptor of surface properties of 3D models. Three of surface properties based descriptors that are mentioned above are combined to generate the hybrid descriptor. Experiment results show that surface properties based hybrid descriptor can get better retrieval effectiveness than any single one.(5). A database with 1,171 colored 3D models is created for evaluation of surface properties based 3D model retrieval due to most of 3D benchmarks can only be used for shape based 3D model retrieval evaluation. In order to evaluate surface properties based 3D model retrieval, a database with colored or textural model and corresponding classifications must be found. This database has 1,171 colored models downloaded from Internet freely. Models in this database are classified according to their functions into 46 classes. Then, models within one class which contains too many models are refined again with surface properties by 59 classes finally. As each model in this database has both shape properties and surface ones, it can be used for evaluation of shape-based 3D model retrieval, surface properties-based 3D model retrieval and combination of them.In conclusion, this thesis dedicates itself to methods of improvements in 3D model retrieval performance and of 3D model retrieval based on surface properties by studying and analyzing previous works. Firstly, by analyzing for different feature vectors extracted from 3D model, a common feature based 3D model retrieval method is proposed. Secondly, by extending Generic Fourier Descriptor to 3D space, a z-axis rotation invariant 3D shape descriptor is proposed. Thirdly, three different surface properties based descriptors, Color Distribution, Spherical Color Distribution and Colored Spherical Harmonic, are proposed. Meanwhile, a hybrid descriptor based on surface properties is proposed by combining three descriptors based on surface properties. Lastly, a database with 1,171 colored 3D models is constructed for evaluation of surface properties based 3D model retrieval. |
