Subspace-based clustering and retrieval of 3-D objects

In this paper, a new method within the framework of principal component analysis is proposed to retrieve 3-D objects. Each object is represented by a set of 2-D images acquired from uniformly distributed viewpoints. The proposed scheme models the distribution of object images using Gaussian clusters. Clusters are created using K-means clustering algorithm. Principal component analysis based nested eigenspaces are obtained for K clusters. The eigenspaces corresponding to each cluster are used for reconfiguring the cluster and for object classification. The signature of every image in the training set is computed by projecting it onto the eigenspaces. The similarity between a query image and signatures is evaluated using the L1 distance measure for the purpose of object recognition. The eigenspaces and signatures constitute the object descriptors. The benchmark datasets such as Princeton Shape Benchmark and Amsterdam Library of Object Images are used to validate the proposed scheme. Experimental comparisons are provided to illustrate the efficacy of the proposed scheme with state-of-the-art 3-D retrieval approaches.