Exploring Non-Significant Areas for Weakly Supervised Semantic Segmentation

This is the authors' abstract. We don't add key points for this paper.

Semantic segmentation is the process of categorizing all pixels in an image. Given the inherent challenges of attaining fine labels, researchers have recently embraced weak labels to mitigate the annotation burden of segmentation. The current work on weakly supervised semantic segmentation (WSSS) mainly focuses on expanding pseudo-label seeds in the salient regions of the image, but there are also many objects outside the salient area that have not been discovered. In this work, we propose an innovative WSSS method by exploring non-significant areas (ENSA). Specifically, we first utilize multiple local views that are randomly clipped from the input image to extract attention. Then, we design a local-global knowledge transfer module (LGKT) for the global network so that the global network can obtain the complementary attention knowledge of multiple local attention maps through online learning to produce high-quality attention maps. To further explore the non-significant areas of complex images, we adopt a NSRM module to generate masked labels. Comprehensive experiments on the PASCAL VOC 2012 dataset illustrate that we achieve state-of-the-art performance compared to existing work.