View-based 3D model retrieval based on distance learning

Abstract

As information technologies develop, 3D model retrieval is paid more and more attentions by researchers. But the limitations of image features poses a great challenge to view-based 3D model retrieval. In this paper, a novel 3D model retrieval method based on distance learning is introduced. The objective function with respective to two latent variables was formulated especially. The variables are the clique information in the original graph and the pairwise clique correspondence constrained by the one-to-one matching. The proposed method has the following benefits: (1) the local and global attributes of a graph with the designed structure can be preserved; (2) redundant and noisy information can be eliminated by strengthening inliers and suppressing outliers; and (3) the difficulty of defining high-order attributes and solving hyper-graph matching can be avoided. By extensive experiments on ETH, NTU60 and MV-RED datasets with Zernike moments, Histograms of Oriented Gradients (HoG) and convolutional neural networks (CNN) features, the effectiveness of the proposed method could be tested.