Abstract

Weakly supervised semantic segmentation (WSSS) methods have received significant attention in recent years, since they can dramatically reduce the annotation costs of fully supervised alternatives. While most previous studies focused on leveraging classification labels, we explore instead the use of image captions, which can be obtained easily from the web and contain richer visual information. Existing methods for this task assigned text snippets to relevant semantic labels by simply matching class names, and then employed a model trained to localize arbitrary text in images to generate pseudo-ground truth segmentation masks. Instead, we propose a dedicated caption processing module to extract structured supervision from captions, consisting of improved relevant object labels, their visual attributes, and additional background categories, all of which are useful for improving segmentation quality. This module uses syntactic structures learned from text data, and semantic relations retrieved from a knowledge database, without requiring additional annotations on the specific image domain, and consequently can be extended immediately to new object categories. We then present a novel localization network, which is trained to localize only these structured labels. This strategy simplifies model design, while focusing training signals on relevant visual information. Finally, we describe a method for leveraging all types of localization maps to obtain high-quality segmentation masks, which are used to train a supervised model. On the challenging MS-COCO dataset, our method moves the state-of-the-art forward significantly for WSSS with image-level supervision by a margin of 7.6% absolute (26.7% relative) mean Intersection-over-Union, achieving 54.5% precision and 50.9% recall.