Research On Content-based Non-rigid Model Retrieval

The rapid development of both the hardware and software makes the usage of3D modelspopular. The usage lies not only in areas like computer aided design, bioscience, medicine,chemistry, industrial and military, but also can be seen in our daily life, such as3D games andmovies. With a growing number of demand of using3D models and the massive appearanceof3D models on the Internet, users may confront a same problem that finding suitable modelsis a difficult job in a short period. Thus, more and more research groups and companiesengaged in this kind of work to find an efficient way to do model retrieval.Firstly, from different aspects, the literature summarized the representative modelretrieval technologies, including rigid and non-rigid ones. This paper mainly focuses onnon-rigid model retrieval, and analyzed the existing drawbacks in current methods. Amongvarious retrieval descriptors, the diffusion-based method is one of the best and efficientsignatures. This kind of method can describe deformable models very well and can resistnoise, topological change and deal with partial models. But the method lacks spatialinformation, thus, it has poor discriminative ability.Secondly, with the problem analyzed above, we came up a new density-invariant modelretrieval strategy. We studied the instable problem for HKS feature matching method usingbags-of-word technology and have improved the precision with our method. In addition, thepaper researched the stability of the precision problem for state-of-the-art and a new multifeature combination retrieval scheme has been proposed. With the new method, we haveimproved the retrieval result significantly.In the proposed combined model retrieval technology, we extracted different types offeatures (local and global). Then, a weighted scheme was used to connect them together. Thecharacteristic of this feature is that it maintains the basic ability and adds new properties formodel description. At last, we stated our method in detail and developed an efficient model retrieval system.We experimented on SHREC’2011dataset and the result shows that the proposed method canreduce redundancy and represent spatial information adequately. As to precision, the methodhas achieved better performance than existing methods.