3D model retrieval using salient visual features from multiple viewpoints

Abstract

In this paper, we propose a view-based 3D shape comparison method that employs local visual features computed at selected “salient” points. The method renders, after normalizing for similarity transformation, images from six canonical viewpoints. Each image is processed by using Scale Invariant Feature Transform (SIFT) algorithm developed by Lowe to detect a set of salient points and to compute a set of feature at each salient point. Our method computes dissimilarity between a pair of 3D models by using these salient points and their respective local features. Our method constrains salient-point correspondence pair by using geometrical proximity, and culls the correspondence pairs to avoid false matches. Experimental evaluation showed that the method achieved the retrieval performance on a par with the best finisher among the SHREC 2006 3D model retrieval contest entrants.