3D Model Retrieval And Its Application

3D model databases will become ubiquitous due to following three facts. First, improved digital multimedia technology, computer vision theory and graphics processing unit have expanded the applications of 3D graphics. Second, popularized modeling tools and 3D scanning devices are making acquisition of 3D models easier and less expensive, creating a large supply of publically available 3D data sets. Finally, the World Wide Web is enabling access to 3D models constructed by people all over the world, providing a mechanism for wide-spread distribution of high quality 3D models.Determining the similarity between 3D shapes is a fundamental task in shape analysis, model classification, model retrieval and shape matching. The work presented in this thesis emphasizes on the research of 3D model representation and similarity definition. Specifically, a global shape representation "Depth Weighted Normal Map" and a local curvature description "Shape Index Histogram" are proposed and applied to 3D model retrieval. The other important aspect is a refined definition of partial similarity between the query pattern and target object.We present an overview of the background, application perspectives and existing problems in chapter 1.Chapter 2 presents a review of current 3D shape analysis technology. We discuss several topics including shape descriptor, similarity measure, relevance feedback, user interface design and system framework. We also give some examples of typical 3D search engine.Chapter 3 proposes a novel 3D model retrieval algorithm based on a new shape descriptor, "Depth Weighted Normal Map"(DWNM), for efficiently matching arbitrary 3D models. By means of uniform orthogonal sampling, we represent the shape signature of each model as a statistical distribution of its surface normals weighted by view-dependent depth. The distribution is further processed by spherical harmonics analysis to construct the final representation. By calculating the distance between shape descriptors of individual 3D models, a faithful similarity measurement is achieved.Chapter 4 and 5 describe our initial solution to partial retrieval of arbitrary 3D models. The task is reformulated as the estimation of similarity transformations be-tween the query pattern and target object. Two stages are thus carried out, which account for the scaling and rotation/translation parts respectively. The scaling is first measured by matching the weighted shape index distributions of both models, while the rotation/translation parts are estimated in Euclidean space. Both are facilitated by means of HMD (earth movers' distance) in searching the correspondence between focused point sets.We conclude the whole thesis in Chapter 6, with a brief discussion of future research directions.