Object Region and Class Learning for Weakly-Supervised Semantic Segmentation

Abstract

In this paper, we study the weakly-supervised semantic segmentation problem that trains the segmentation network using image-level label. In conventional WSSS approaches, to train the segmentation network in the absence of the pixel-level labels, the classification network is used to generate the pseudo-label. The potential problems in the conventional approaches are: 1) the labeled regions in pseudo-label is small and sparse, 2) the pseudo-labels for complex images are inaccurate. To handle this, we propose a novel WSSS framework that can train the segmentation network without generating the pseudo-label. By masking the input image using the segmented output and delivering the masked image to the classification network, the segmentation network is penalized if the segmented output is inaccurate. We show that the proposed approach can effectively train the segmentation network.