A Reconfigurable 3D Engineering Shape Search System: Part II — Database Indexing, Retrieval, and Clustering

Abstract

This paper introduces database and related techniques for a reconfigurable, intelligent 3D engineering shape search system, which retrieves similar 3D models based on their shape content. Feature vectors, which are numeric “fingerprints” of 3D models, and skeletal graphs, which are the “minimal representations of the shape content” of a 3D model, represent the shape content. The Euclidean distance of the feature vectors, as well as the distance between skeletal graphs, provides indirect measures of shape similarity between the 3D models. Critical database issues regarding 3D shape search systems are discussed: (a) database indexing, (b) semantic gap, (c) subjectivity of similarity, and (d) database clustering. An Rtree based multidimensional index is used to speed up the feature-vector based search operation, while a decision treebased approach is used for efficiently indexing/searching skeletal graphs. Interactions among users and the search system, such as relevance feedback and feature vector reconfiguration, are used to bridge the semantic gap and to customize the system for different users. Database clustering of the R-tree index is compared with that generated by a selforganizing map (SOM). Synthetic databases and real 3D model databases are employed to investigate the efficiency of the multidimensional index and the effectiveness of relevance feedback.

Relevance Feedback in Content-based 3D Object Retrieval A Comparative Study

“…In the hybrid scheme, LDA was used when the user's relevance feedback was limited since it was found to perform better than BDA in that case, while BDA performed better for larger numbers of labeled results. In the work of Lou et al [13], the system moves the query point to the centroid of the relevant results and re-weights the feature dimensions based on their contribution in discriminating the relevant from the irrelevant results, by measuring the features variance. In the approach of Atmosukarto et al [2], the system learned the importance of different feature representations and pushed possible irrelevant models to the bottom of the retrieval list.…”

“…Relevance feedback was first used to improve text retrieval [19], later on successfully employed in image retrieval systems [6], [9], [10], [15], [18], [20] and lately in 3D object retrieval systems [1], [3], [8], [13]. It is the information of relevance with respect to a subset of the retrieved results, acquired from the user's interaction with the retrieval system.…”

PANORAMA: A 3D Shape Descriptor Based on Panoramic Views for Unsupervised 3D Object Retrieval

“…Apart from a few studies that incorporated RF in 3D object retrieval [42], [43], [44], [45], [46], [47], [48], LRF has only lately been examined in [5].…”

A survey of content based 3D shape retrieval methods

“…Iyer et al [39,51] use both global features and skeletal graphs to describe volume models, obtained by voxelizing solid models. They obtain a skeletal graph by a thinning algorithm iteratively eroding voxels until a one-voxel width skeleton is left.…”