A New Distance Measure for Non-rigid Image Matching

  • Authors:

  • Anand Rangarajan;Haili Chui;Eric Mjolsness

  • Affiliations:

  • -;-;-

  • Venue:
  • EMMCVPR '99 Proceedings of the Second International Workshop on Energy Minimization Methods in Computer Vision and Pattern Recognition
  • Year:
  • 1999

Quantified Score

Hi-index 0.00

Visualization

Abstract

We construct probabilistic generative models for the non-rigid matching of point-sets. Our formulation is explicitly Platonist. Beginning with a Platonist super point-set, we derive real-world point-sets through the application of four operations: i) spline-based warping, ii) addition of noise, iii) point removal and iii) amnesia regarding the pointto-point correspondences between the real-world point-sets and the Platonist source. Given this generative model, we are able to derive new non-quadratic distance measures w.r.t. the "forgotten" correspondences by a) eliminating the spline parameters from the generative model and by b) integrating out the Platonist super point-set. The result is a new non-quadratic distance measure which has the interpretation of weighted graph matching. The graphs are related in a straightfoward manner to the spline kernel used for non-rigid warping. Experimentally, we show that the new distance measure outperforms the conventional quadratic assignment distance measure when both distances use the same weighted graphs derived from the spline kernel.