Résumé: Reliable curve matching is a difficult yet important problem in many vision-based applications including image-based modeling. We describe in this paper two aspects of our research in this area: a new algorithm for curve matching (including lines) within a probabilistic relaxation framework, and an approach of incorporating previously matched points/corners to guide curve matching. We propose similarity-invariant unary and binary measurements suitable for curves, and introduce an additional measurement to model the uncertainty of the binary measurements. The uncertainty measure is proven to be very important in computing the matching support from neighboring matches. We also show how to use a set of previously matched points/corners to guide the curve matching. The role of the corner guidance is explicitly modeled by a set of unary measurements and a similarity function under the same relaxation framework. Preprocessing techniques contributing to the success of our curve matching techniques are also developed and discussed. Experiments with complex real scenes show that the rate of correct matching is higher than 98%.

Langue: Anglais