Maximum Weight Clique Based Image Object Co-localization With Mutex Constraint

With the increasingly abundant of images shared on the internet,the localizing and retrieving of generic object become growing demand.Given a set of images containing objects from the same category,the task is to identify and localize each instance in each image.We can treat the object localization in each image as a independent process using traditional method but has several shortcomings:on the one hand,it may ignore the common visual patterns in the set of images to be detected.On the other hand,it is necessary to provide an empirical model of the object or a large number of manually labeled training sets,with a time-consuming pre training process.In recent years,with the gradual development of the idea of "Joint",the Co-localization of similar objects directly from a group of images is widely concerned.In order to find similar objects from the batch image,we proposed a Maximum Weight Cliques based object co-localization with mutex constraint.This method formulated the task of co-localization as finding the Maximum Weight Cliques in a weighted graph,in which the set of nodes decribing the candidates get from objectness with their prior and the edges representing the similarity between pair of vertexs.In the aspect of priori acquisition,a new method based on nearest neighbor distribution and saliency is proposed.In addition,a mutex constraint binary matrix modeled on overlap and aspect ratio is added to connect the nodes in the graph,which can be used to limit selecting overlaping objects or candidates with large difference in shape.After solving the optimization problem,this method is able to co-locate the objects of the same class across a set of distinct images in unsupervised way.Then,we extend the method in video co-localization.We perform an extensive evaluation of our method compared to previous state-of-the art approaches on the challenging PASCAL VOC 2007 and Object Discovery datasets,and got good result.