From region based image representation to object discovery and recognition

This paper presents an overview of the work we have done over the last several years on object recognition in images from region-based image representation. The overview focuses on the following related problems: (1) discovery of a single 2D object category frequently occurring in a given image set; (2) learning a model of the discovered category in terms of its photometric, geometric, and structural properties; and (3) detection and segmentation of objects from the category in new images. Images in the given set are segmented, and then each image is represented by a region graph that captures hierarchy and neighbor relations among image regions. The region graphs are matched to extract the maximally matching subgraphs, which are interpreted as instances of the discovered category. A graph-union of the matching subgraphs is taken as a model of the category. Matching the category model to the region graph of a new image yields joint object detection and segmentation. The paper argues that using a hierarchy of image regions and their neighbor relations offers a number of advantages in solving (1)-(3), over the more commonly used point and edge features. Experimental results, also reviewed in this paper, support the above claims. Details of our methods as well of comparisons with other methods are omitted here, and can be found in the indicated references.