3D Shape Recognition using Sparse Representation Errors of Local Descriptors

Abstract

In this paper, a new algorithm is proposed for 3D shape recognition using sparse representation of local descriptors. We use weighted sparse representation errors of local descriptors using various class dictionaries to obtain posterior probability of belonging a local descriptor to a shape class. Then the maximum likelihood approach is used to classify local descriptors. We utilize a majority voting algorithm to recognize 3D shapes by using the classification results of local descriptors. In contrast to existing approaches that use the geometrical properties of local points to extract salient points, we use a new method to extract salient points using sparse representation. We define a point as a salient point if it can be correctly classified using the proposed algorithm. We also use a new approach for feature weighting. The experimental results as well as the comparison of the proposed algorithm with a state-of-the-art approach show the promise of the proposed shape recognition algorithm.