In this paper, we propose a dynamic multi-descriptor fusion （DMDF） approach to improving the retrieval accuracy of 3-dimensional （3D） model retrieval systems. First, an independent retrieval list is generated by u...In this paper, we propose a dynamic multi-descriptor fusion （DMDF） approach to improving the retrieval accuracy of 3-dimensional （3D） model retrieval systems. First, an independent retrieval list is generated by using each individual descriptor. Second, we propose an automatic relevant/irrelevant models selection （ARMS） approach to selecting the relevant and irrelevant 3D models automatically without any user interaction. A weighted distance, in which the weight associated with each individual descriptor is learnt by using the selected relevant and irrelevant models, is used to measure the similarity between two 3D models. Furthermore, a descriptor-dependent adaptive query point movement （AQPM） approach is employed to update every feature vector. This set of new feature vectors is used to index 3D models in the next search process. Four 3D model databases are used to compare the retrieval accuracy of our proposed DMDF approach with several descriptors as well as some well-known information fusion methods. Experimental results have shown that our proposed DMDF approach provides a promising retrieval result and always yields the best retrieval accuracy.展开更多
In order to avoid the discussion of equation （1.1）, this paper employs the proof method of Liang （2012） to consider the re-weighted Nadaraya-Watson estimation of conditional density. The established results genera...In order to avoid the discussion of equation （1.1）, this paper employs the proof method of Liang （2012） to consider the re-weighted Nadaraya-Watson estimation of conditional density. The established results generalize those of De Gooijer and Zerom （2003）. In addition, this paper improves the bandwidth condition of Liang （2012）.展开更多
基金supported in part by“MOST”under Grants No.102-2632-E-216-001-MY3 and No.104-2221-E-216-010-MY2
文摘In this paper, we propose a dynamic multi-descriptor fusion （DMDF） approach to improving the retrieval accuracy of 3-dimensional （3D） model retrieval systems. First, an independent retrieval list is generated by using each individual descriptor. Second, we propose an automatic relevant/irrelevant models selection （ARMS） approach to selecting the relevant and irrelevant 3D models automatically without any user interaction. A weighted distance, in which the weight associated with each individual descriptor is learnt by using the selected relevant and irrelevant models, is used to measure the similarity between two 3D models. Furthermore, a descriptor-dependent adaptive query point movement （AQPM） approach is employed to update every feature vector. This set of new feature vectors is used to index 3D models in the next search process. Four 3D model databases are used to compare the retrieval accuracy of our proposed DMDF approach with several descriptors as well as some well-known information fusion methods. Experimental results have shown that our proposed DMDF approach provides a promising retrieval result and always yields the best retrieval accuracy.
基金supported by National Natural Science Foundation of China(No.11301084)Natural Science Foundation of Fujian Province(No.2014J01010)
文摘In order to avoid the discussion of equation （1.1）, this paper employs the proof method of Liang （2012） to consider the re-weighted Nadaraya-Watson estimation of conditional density. The established results generalize those of De Gooijer and Zerom （2003）. In addition, this paper improves the bandwidth condition of Liang （2012）.
