3D Model Retrieval System Based On BoW-SIFT Model And Hierarchical Mesh Feature

This article aims to study a 3D model retrieval system with a better performance. The model retrieval system is a relatively new but difficult research area, which has very broad prospects for future applications. BoW(Bag of Words) technology originated in document retrieval systems and got a great success. SIFT feature extraction technique was proposed by the D.G.Lowe 1999, concluded in 2004, and obtained good performance in the image retrieval area. In this paper, we applied them to the model retrieval system. Experiments showed that the method based on BoW-SIFT model (BoW and SIFT feature extraction) was feasible and efficient.The content of this paper is as follows.1. We focused on the principles of the BoW-SIFT model, and proposed using BoW-SIFT model to implement a three-dimensional model retrieval system. As the lack of mesh structure and semantic information in BoW, one of mesh segmentations (2-way hierarchical segmentation) was introduced to our system, and the extracted mesh features after mesh segmention (such as the number of regions, the regional average geodesic distance, the center of the model) was proposed in BoW-SIFT system.2. A model retrieval system based on BoW-SIFT model was designed and carried out. The system was based on the Princeton University 3D model library. The system could work well after the processes of view generation, SIFT feature extraction, KMeans clustering (generating words bags), BoW representation, similarity measurement, visualization of search results, evaluation of search and so on.3. The parameters of the system was determined by some experiments, such as the dictionary size, the selection of the initial seed point before clustering, the number of views and the selection of distance criteria and so on. Compared to other 3D retrieval systems (such as LFD, D2 distance), the results showed that the retrieval performance of our system was relatively better than others. Another important result was that BoW-SIFT system using mesh features also had better performance than BoW-SIFT system without mesh features.