Learning to Associate Image Features with CRF-Matching
Proc. of International Symposium on Experimental Robotics (ISER), 2008
Abstract
This paper presents a supervised learning algorithm for image feature
matching. The algorithm is based on Conditional Random Fields which provides a
mechanism for globally reason about the associations. The novelty of this work
is twofold: (i) the use of Delaunay triangulation as the graph structure for a
probabilis- tic network to reason about image feature association; (ii) the
combination of local and joint features to enforce consistency in a
theoretically sound statistical learning procedure. Experimental results show
that our approach outperforms RANSAC in our challenging datasets consisting of
indoor and outdoor images, with significant occlusion, blurring, rotational and
translational transformations.