Recognition with Local Features: the Kernel Recipe

  • Authors:
  • Christian Wallraven;Barbara Caputo;Arnulf Graf

  • Affiliations:
  • -;-;-

  • Venue:
  • ICCV '03 Proceedings of the Ninth IEEE International Conference on Computer Vision - Volume 2
  • Year:
  • 2003

Quantified Score

Hi-index 0.00

Visualization

Abstract

Recent developments in computer vision have shown that localfeatures can provide efficient representations suitable for robustobject recognition. Support Vector Machines have been establishedas powerful learning algorithms with good generalizationcapabilities. In this paper, we combine these two approaches andpropose a general kernel method for recognition with localfeatures. We show that the proposed kernel satisfies the Mercercondition and that it is suitable for many established localfeature frameworks. Large-scale recognition results are presentedon three different databases, which demonstrate that SVMs with theproposed kernel perform better than standard matching techniques onlocal features. In addition, experiments on noisy and occludedimages show that local feature representations significantlyoutperform global approaches.