Categorization via Agglomerative Correspondence Clustering

This paper presents computationally efficient object detection, matching and categorization via Agglomerative Correspondence Clustering (ACC). We implement ACC for feature correspondence and object-based image matching exploiting both photometric similarity and geometric consistency from local invariant features. Objectbased image matching is formulated here as an unsupervised multi-class clustering problem on a set of candidate feature matches linking maximally stable external regions features and scale invariant features in the framework of hierarchical agglomerative clustering. The algorithm is capable to handle significant amount of outliers and deformations such as scaling and rotation as well as multiple clusters, thus powering simultaneous feature matching and clustering from real-world image pairs with significant clutter and multiple objects. The experimental assessment on feature correspondence, object recognition, and objectbased image matching demonstrates that, this method is capable of rigorously handling scaling, rotation, and deformation and can be applied to a wide range of image matching and object recognition and categorization related real-world problems.