EE SEminar: Inference in Perturbation Models for Enhanced Matching

~~
Speaker: Chen Harel
M.Sc. student under the supervision of Prof. Daniel Cohen-Or and Prof. Shai Avidan

Wednesday, June 15th, 2016 at 15:30
Room 011, Kitot Bldg., Faculty of Engineering

Inference in Perturbation Models for Enhanced Matching

Matching meaningful correspondences is a challenging and important problem in computer vision. In particular, bipartite graph matching is a useful tool in many visual applications. In this work, we introduce the notion of perturbations for enhanced matching, and present the power of perturbation matching in the context of shape matching of image contours.
Perturbation models are high dimensional probability models that measure the stability of max-prediction to random shifts of the computed scores. The probability of a given matching is the volume of scores for which this matching is the highest scoring matching.
We demonstrate the effectiveness of our approach in pair of images with background clutter, showing that inference with perturbation models is more robust than finding the highest scoring matching. We also show that for perturbation matching it is favorable to use logistic perturbation models rather than Gumbel perturbation models, proving that inference using logistic perturbation models is more efficient.