Learning Distance Functions for Automatic Annotation of Images

  • Authors:
  • Josip Krapac;Frédéric Jurie

  • Affiliations:
  • INRIA Rhône-Alpes, Saint Ismier Cedex, France 38334;INRIA Rhône-Alpes, Saint Ismier Cedex, France 38334

  • Venue:
  • Adaptive Multimedial Retrieval: Retrieval, User, and Semantics
  • Year:
  • 2007

Quantified Score

Hi-index 0.00

Visualization

Abstract

This paper gives an overview of recent approaches towards image representation and image similarity computation for content-based image retrieval and automatic image annotation (category tagging). Additionaly, a new similarity function between an image and an object class is proposed. This similarity function combines various aspects of object class appearance through use of representative images of the class. Similarity to a representative image is determined by weighting local image similarities, where weights are learned from training image pairs, labeled "same" and "different", using linear SVM. The proposed approach is validated on a challenging dataset where it performed favorably.