Abstract
View-based 3D model retrieval is an important and challenging task in computer vision, which can be utilized in many applications such as landmark detection, image set classification, etc. Representation view selection and similarity measure are two key problem in view-based 3D model retrieval. Many classic algorithms were proposed to handle these two problems. However, they were often independent to consider these two problems while ignoring the contact with each other. In this paper, we proposed a joint subgraph learning & matching method (SGLM) via Markov Chain Monte Carlo (MCMC) to handle view-based 3D model retrieval problem, which effectively combine representation view extraction with similarity measure process to find the best matching result. The proposed (SGLM) can benefit: 1) considering the correlation between representation view selection and similarity measure, which can effectively improve the final performance of retrieval; 2) eliminating redundant visual information by subgraph learning; 3) learning representation views automaticly in similarity measure process. We validate the SGLM based on 3D model retrieval on ETH, PSB, NTU and MVRED datasets. Extensive comparison experiments demonstrate the superiority of the proposed method.
Highlights
Compared with 2D images, 3D models have the advantages to reappear the objects’spatial and structure information which is more suitable for the visual perception system of humans
In order to handle these problems, we proposed a novel sub-graph learning matching method (SGLM) based on Markov Chain Monte Carlo framework, like in Fig.1, to handle 3D model retrieval problem, which can combine the representation views selection with similarity measure steps into one step and effectively improve the effectiveness of retrieval
In the ETH, the proposed method can achieve a gain of 4% - 15%, 1% - 13%, 1% - 3%, 1% - 3%, 1% - 14% in terms of Nearest Neighbor (NN), First Tier (FT), Second Tier (ST), F-measure, Discounted Cumulative Gain (DCG), and achieve a decline of 3%-27% in terms of Average Normalized Modified Retrieval Rank (ANMRR)
Summary
Compared with 2D images, 3D models have the advantages to reappear the objects’spatial and structure information which is more suitable for the visual perception system of humans. Representation views selection: many recent methods often apply some simple view clustering and center selection via visual low-level features These methods are not robust and effective enough to filter redundant and noisy data. Similarity measurement: statistical model, probability model are utilized to handle similarity measure between two different 3D models These methods all focus on the distribution of visual feature and ignore the spatial and structure information of views. In order to handle these problems, we proposed a novel sub-graph learning matching method (SGLM) based on Markov Chain Monte Carlo framework, like in Fig., to handle 3D model retrieval problem, which can combine the representation views selection with similarity measure steps into one step and effectively improve the effectiveness of retrieval.
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
