Research On Feature Extraction Method Of Non-rigid3D Model

With the development of computer graphics and related technology,3D models areplaying an important role in many areas, such as mechanical CAD, computer vision, virtualreality, movies, games, medicine and so on. More and more3D model databases have beencreated due to the explosive growth of the number of models. How to find the target modelquickly and exactly has become an urgent problem to be solved. The computer vision andpattern recognition communities have recently witnessed a surge of feature-based methods inimage and video retrieval applications. Recently content-based3D model retrieval hasbecome a hot research topic in recent years learned by many scholars.3D models consist of rigid models and non-rigid models. It has achieved prominenteffect on the feature extraction and retrieval of the rigid model in recent years. Featuredescription and retrieval of non-rigid model face more complex problems due to the isometrytransformation of itself. This paper summarizes the various feature descriptors for thenon-rigid model, while takes analysis and elaboration of important issue existing in thenon-rigid3D model retrieval.When select or design a feature descriptor we should focus on two aspects: First, whatkind of shape properties to describe of the feature descriptor; Second, what kind of shapetransformations to be robust of the feature descriptor. Good feature descriptor can fullycharacterize the intrinsic properties of the non-rigid model while remain robust to isometrictransformation of the model. Heat kernel signature and wave kernel signature based onlaplace operator spectral decomposition can fully describe the intrinsic properties of thenon-rigid model while can be robust to the non-rigid transformation such as shape distortionsand postural changes and so on. We apply the wave kernel signature to describe thecharacteristics of models and introduce the bag of words model to quantify the characteristics,the bag of words model which derives from the text retrieval. The method based on the bag ofwords model fails to describe the spatial distribution of feature, therefore, we introduce the hierarchical combination matching into the feature compare, and build a map between thedivided regions of two models, and then achieve accurate feature matching based on patch bypatch, which successfully introduce the spatial information into feature matching. Verified byexperiment, the3D model retrieval method proposed in this paper based on the wave kernelsignature and hierarchical combination matching can make sure more accurate featurematching, so as to enhance the precision of retrieval.