Comparison of Dimension Reduction Methods for Database-Adaptive 3D Model Retrieval
Distance measures, along with shape features, are the most critical components in a shape-based 3D model retrieval system. Given a shape feature, an optimal distance measure will vary per query, per user, or per database. No single, fixed distance measure
- PDF / 970,811 Bytes
- 15 Pages / 430 x 660 pts Page_size
- 22 Downloads / 157 Views
Abstract. Distance measures, along with shape features, are the most critical components in a shape-based 3D model retrieval system. Given a shape feature, an optimal distance measure will vary per query, per user, or per database. No single, fixed distance measure would be satisfactory all the time. This paper focuses on a method to adapt distance measure to the database to be queried by using learning-based dimension reduction algorithms. We experimentally compare six such dimension reduction algorithms, both linear and non-linear, for their efficacy in the context of shape-based 3D model retrieval. We tested the efficacy of these methods by applying them to five global shape features. Among the dimension reduction methods we tested, non-linear manifold learning algorithms performed better than the other, e.g. linear algorithms such as principal component analysis. Performance of the best performing combination is roughly the same as the top finisher in the SHREC 2006 contest.
1 Introduction Research on shape-based retrieval of 3D models [24, 13, 26] has recently gained attention. A shape-based 3D model retrieval system retrieves, given a query, a set of shape models ranked by their shape-based similarity to the query. The query may be texts, 2D sketches, 3D sketches, or 3D models. In this paper, we assume the query is a 3D model defined, for example, as a set of polygons and that the system retrieves 3D models similar in their shape to the query. Two of the most significant technical challenges for shape-based retrieval of 3D models are feature extraction and distance computation. We first have to device a compact yet expressive shape feature that can be extracted and compared with reasonable computational cost. In feature extraction, compatibility of the shape feature with various shape representations, e.g., polygon soup and voxel enumeration solid, is an important factor to be considered. Then, distance, or dissimilarity among a pair of models to be compared must be computed. It is desirable that the feature and distance measure be adaptive to a database, to a user, or even to a specific query. For example, a combination of feature and distance metric that works well for comparing human face models may be sub-optimal for comparing screws. Or, the bunny model I wanted yesterday may be different from a bunny model I want today. So-called “curse of N. Boujemaa, M. Detyniecki, and A. Nürnberger (Eds.): AMR 2007, LNCS 4918, pp. 196–210, 2008. © Springer-Verlag Berlin Heidelberg 2008
Comparison of Dimension Reduction Methods
197
dimensionality”, which discourages higher dimensional shape feature vectors, also comes into play as the dimensionality of 3D shape feature tend to be quite high. In our previous work [20], we adopted Xhaofei He’s approach [12] for a significant gain in retrieval performance. Instead of the distance in the original feature (or input) space, the method uses geodesic distance on a subspace, or a manifold, spanned by the features of 3D models in a database. He used the Laplacian Eigenmaps
Data Loading...
