Densifying Distance Spaces for Shape and Image Retrieval

  • Authors:

  • Xingwei Yang;Xiang Bai;Suzan Köknar-Tezel;Longin Jan Latecki

  • Affiliations:

  • Image Analytics Lab, GE Global Research, Niskayuna, USA 12309;Dept. of Electronics and Information Engineering, Huazhong University of Science and Technology (HUST), Wuhan, P.R. China 430074;Department of Computer and Information Sciences, Temple University, Philadelphia, USA 19122;Department of Computer and Information Sciences, Temple University, Philadelphia, USA 19122

  • Venue:
  • Journal of Mathematical Imaging and Vision
  • Year:
  • 2013

Quantified Score

Hi-index 0.00

Visualization

Abstract

Sparse data sets are an ever-present problem in many fields of computer science. In the shape retrieval community, several researchers use graph transduction algorithms to reveal the underlying structure of the shape manifold. Without an infinite number of shapes, the data sets can only imprecisely describe the shape manifold. For this problem, adding synthetic data points can be very effective. However existing methods add synthetic points only in feature space. In distance spaces, which are often non-metric and are widely used in bioinformatics, time series classification, shape similarity, and other domains, it is impossible to use these standard, feature-based methods, such as SMOTE, to insert synthetic points. Instead, we present an innovative approach that adds synthetic points directly to distance spaces. We call these synthetic points ghost points since they are not represented by vectors of features, and consequently, cannot be directly visualized. However, we can define the distances of ghost points to all other data points. Our experimental results on standard data sets show that ghost points not only significantly improve the accuracy of shape retrieval, but also the accuracy of image retrieval. We also discuss the conditions that allow the ghost points to improve retrieval results.