Collaborative Semantic Segmentation with Image Labels

Abstract

Weakly-supervised semantic segmentation has recently received much attention since it needs less fine-grained annotations than fully-supervised learning. Most existing studies use attention maps from the classification network as supervision, which suffers from only locating small discriminative parts of objects and lacking precise boundaries. In this study, we propose a collaborative segmentation network consisting of a localization sub-network and a segmentation sub-network. The localization sub-network takes only image-level labels as supervision. The pseudo masks generated by the segmentation sub-network and the localization sub-network, providing rough localization information of objects and the shape information respectively, are mixed-up adaptively and used as supervision for the segmentation sub-network. The two sub-networks share the same backbone and are shown to mutually enhance each other during training. We evaluate the proposed method on PASCAL VOC 2012 Semantic Segmentation benchmark and achieve new state-of-the-art.