Object Verification via Mining Visual Patterns

Abstract

In this paper we propose a novel object verification method based on mining local visual patterns lying in a image pair. Our method is robust to object deformation because (i) any number of local visual patterns can be mined independently of their locations and (ii) each pattern is extracted via clustering of the correspondence graph that embeds geometric similarities of keypoint quadruplets. We validate our approach in an experiment of detecting deformed planar objects in a noisy image collection.