View-based clustering and content-based retrieval of 3-D objects from their appearances

This paper addresses the problem of content-based image retrieval (CBIR) scheme of 3-D objects using 2-D views. A clustering algorithm has been proposed to group the images in from low-dimensional subspaces of the data set. Each cluster has images with distinct range of viewpoints. Discrete clusters are generated in a tree-like structure. The training set of the retrieval scheme is formed using the collection of clusters obtained. Content-based image retrieval is carried out in two stages. The quality of clusters obtained by the proposed clustering scheme is compared with state-of-the-art K-means algorithm and the results validate the superiority of the proposed clustering scheme. The benchmark data set for objects with large pose variations has been used to illustrate the effectiveness of proposed grouping scheme and the retrieval process.