Research On View-based 3D Model Retrieval

The development of information technology drives the rapid growth of 3D technologies,which makes 3D technologies gradually permeate in every corner of people's life,such as 3D movies,3D body feeling game,3D imaging,3D medical technologies and so on.In the past,people could only get information through some 2D images and video from the Internet.But now,people can watch 3D information anytime and anywhere through the virtual reality technology.With the development of those technologies,the utilization of 3D model technologies is becoming more and more frequent,which makes the number of 3D models gradually tends to develop in the direction of big data.Therefore,3D model retrieval technology arises at the historic moment reasoning that how to achieve an accurate retrieval which meets the needs of users is the focus of current research.Characters were used to describe the 3D model in the early 3D model retrieval field which had a strong subjectivity.Now 3D model retrieval methods based on content where the similarity calculate between two models is based on the features extracted from the models are most widely used.In this paper,we focus on the view-based 3D model retrieval and propose two retrieval methods on the basis of them: 1)Supervised Multi-view Feature Learning 3D Object Retrieval: the multiple views of 3D models are regarded as the input of retrieval and singular value decomposition algorithm are used to get the feature subspace.The similarity is calculated based on the feature dimension reduction equation in the subspace;2)Multi-view Feature Mapping 3D Object Retrieval: the multiple views and single view of 3D models are regarded as the input of retrieval and the feature transformation matrix is gained according to the multiple-view Gaussian kernel,the single-view Gaussian kernel and the mixed-view Gaussian kernel.The similarity is calculated after the features of multiple views and single view features are transformed into the same space.In this paper,we utilize 3D model datasets such as Eidgen?ssische Technische Hochschule,National Taiwan University,Multi-view RGB-D Object Dataset and Princeton shape benchmark and view features such as Zernike moments,Histograms of Oriented Gradients and Convolutional Neural Network.The results are compared with several classic algorithms and the massive experimental results show the effectiveness of the proposed methods.