Distance metric learning and feature combination for shape-based 3D model retrieval

This paper proposes a 3D model retrieval algorithm that employs an unsupervised distance metric learning with a combination of appearance-based features; two sets of local visual features and a set of global features. These visual features are extracted from range images rendered from multiple viewpoints about the 3D model to be compared. The local visual features are bag-of-features histograms of a set of Scale Invariant Feature Transform (SIFT) features by Lowe [7] sampled at either salient or dense and random points. The global visual feature is also a SIFT feature sampled at an image center. The proposed method then uses an unsupervised distance metric learning based on the Manifold Ranking (MR) [15] to compute distances between these features. However, the original MR may not be effective when applied to a set of features having certain distance distribution. We propose an empirical method to adjust the distance profile so that the MR becomes effective. Experiments showed that the retrieval algorithm using a linear combination of distances computed from the proposed set of features by using the modified MR performed well across multiple benchmarks having different characteristics.