Reinforcement of keypoint matching by co-segmentation in object retrieval: Face recognition case study

Abstract

The paper investigates a certain group of problems in visual detection and identification of near-identical the-same-class objects. We focus on difficult problems for which: (1) the intra-class visual differences are comparable to inter-class differences, (2) views of the objects are distorted both photometrically and geometrically, and (3) objects are randomly placed in images of unpredictable contents. Since detection of the-same-person faces in complex images is one of such problems, we use it as the case study for the proposed approach. The approach combines a relatively inexpensive technique of near-duplicate fragment detection with a novel co-segmentation algorithm. Thus, the initial pool of matching candidates can be found quickly (however, with limited precision, i.e. many false positives can be detected). It is shown that the subsequent co-segmentation can effectively reject false positives and accurately extract the matching objects from random backgrounds. © 2012 Springer-Verlag.