Hypergraph based feature fusion for 3-D object retrieval

In view-based 3D object retrieval, each object is represented by a set of image views. 3D object retrieval becomes a group matching problem under such definition. Recent works have shown the effectiveness of hypergraph learning that computes the distance between 3D objects by solving a hypergraph structure problem. However, the single feature used in most of state-of-the-art works is often not sufficient to describe a 3D object. In this paper, we propose a feature fusion method based on hypergraph for 3D object retrieval. Besides the frequently used Zernike moments feature, we propose a Dense Kernel Local Binary Feature (DKLBP) feature for 3D object view description. A feature fusion method is proposed under the hypgraph framework. Experiments are conducted on the popular ETH-80 and National Taiwan University 3D model datatsets. Extensive experimental results show that the proposed approach has made significant performance improvement compared to other competitive approaches in recent works.