3D Model Retrieval Using Density-Based Silhouette Descriptor

Fast and Accurate scanning technology equipped with shape modelingand rendering tools and the development of internet have enabled the meansof acquiring, designing, and manipulating complete3D models of real-worldobjects. With growing interest of3D models, their effective retrieval fromlarge databases is becoming a hot spot. Text-based systems, much like othermedia application, would remain severely limited in describing andretrieving3D models. Content-based systems, on the other hand, generateshape descriptors by using models’ own information.Among many content-based retrieval algorithms, density-based framework (DBF) and silhouette descriptor (SIL) have high retrieval accuracyover a broad and heterogeneous set of shape categories. DBF describes3Dobjects with multivariate probability density functions of chosen shapefeatures. SIL compares the similarity of two models by the similarity of their2D silhouettes. We address content-based retrieval of complete3D objectsby introducing density-based framework into silhouette descriptor. We callthe proposed shape descriptor as density-based silhouette descriptor (DBS)in that it uses probability density functions to describe the featuredistributions of a given3D object’s multivariate2D silhouettes’ features.DBS is more insensitive to mesh resolution than DBF and it providesmore precise description of the shape of silhouettes than SIL. Also DBSintroduces a permutation property which can be used to guarantee invarianceat the shape matching stage. As proven by extensive retrieval experiments onseveral3D databases, DBS provides state-of-the-art discrimination over abroad and heterogeneous set of shape categories.