Unsupervised Learning of Foreground Object Segmentation

Abstract

Unsupervised learning poses one of the most difficult challenges in computer\nvision today. The task has an immense practical value with many applications in\nartificial intelligence and emerging technologies, as large quantities of\nunlabeled videos can be collected at relatively low cost. In this paper, we\naddress the unsupervised learning problem in the context of detecting the main\nforeground objects in single images. We train a student deep network to predict\nthe output of a teacher pathway that performs unsupervised object discovery in\nvideos or large image collections. Our approach is different from published\nmethods on unsupervised object discovery. We move the unsupervised learning\nphase during training time, then at test time we apply the standard\nfeed-forward processing along the student pathway. This strategy has the\nbenefit of allowing increased generalization possibilities during training,\nwhile remaining fast at testing. Our unsupervised learning algorithm can run\nover several generations of student-teacher training. Thus, a group of student\nnetworks trained in the first generation collectively create the teacher at the\nnext generation. In experiments our method achieves top results on three\ncurrent datasets for object discovery in video, unsupervised image segmentation\nand saliency detection. At test time the proposed system is fast, being one to\ntwo orders of magnitude faster than published unsupervised methods.\n