Efficient Object Region Discovery for Weakly-supervised Semantic Segmentation

Abstract

Deep Convolutional Neural Networks (DCNNs) provide the leading performance in the semantic segmentation task. However, collecting large-scale pixel-level annotations for training such a DCNN is labor intensive and not cost-effective. In this paper, we propose a small to large (STL) Filed-Of-View framework to train semantic segmentation networks from image level annotations. Specifically, we first train a small Filed-Of-View segmentation network (SFN) with the image-level annotations to discover initial object regions effectively. These localized regions are then combined with saliency maps to construct hypotheses on pixel-level annotations, using which a large Filed-Of-View segmentation network (LFN) is learned. To further enhance the segmentation quality, the object regions generated by LFN are verified with saliency maps, and thus the hypotheses are refined in an iterative manner. The converged hypotheses serve as the supervision information to learn a more powerful LFN for semantic segmentation. Extensive experimental results on PASCAL VOC 2012 segmentation benchmark well demonstrates the superiority of our proposed methods compared with the state-of-the-art.