Unsupervised geodesic convex combination of shape dissimilarity measures

Bilal Mokhtari, Kamal Eddine Melkemi, Dominique Michelucci, Sebti Foufou

Research output: Contribution to journalArticlepeer-review

Abstract

Dissimilarity or distance metrics are the cornerstone of shape matching and retrieval algorithms. As there is no unique dissimilarity measure that gives good performances in all possible configurations, these metrics are usually combined to provide reliable results. In this paper we propose to compute the best linear convex, or weighted, combination of any set of measured shape distances to enhance shape matching algorithms. To do so, a database is represented as a graph, where nodes are shapes and the edges carry the convex combination of dissimilarity measures. Weights are chosen to maximize the weighted distances between the query shape and shapes in the database. The optimal weights are solutions of a linear programming problem. This fully unsupervised method improves the outcomes of any set of shape similarity measures as shown in our experimental results performed on several popular 3D shape benchmarks.

Original languageEnglish (US)
Pages (from-to)46-52
Number of pages7
JournalPattern Recognition Letters
Volume98
DOIs
StatePublished - Oct 15 2017

Keywords

  • Convex combination
  • Descriptors
  • Dissimilarity measures
  • Fusion of descriptors
  • Geodesic distance
  • Shape matching

ASJC Scopus subject areas

  • Software
  • Signal Processing
  • Computer Vision and Pattern Recognition
  • Artificial Intelligence

Fingerprint

Dive into the research topics of 'Unsupervised geodesic convex combination of shape dissimilarity measures'. Together they form a unique fingerprint.

Cite this