Coupling Global Context and Local Contents for Weakly-Supervised Semantic Segmentation

• DOI: 10.48550/arxiv.2304.09059
• Publication Date: 2023-01-01
• Authors: Chunyan Wang; Dong Zhang; Liyan Zhang; Jinhui Tang
• Keywords: Pascal (unit); Feature (linguistics)

Thanks to the advantages of friendly annotations and satisfactory performance, Weakly-Supervised Semantic Segmentation (WSSS) approaches have been extensively studied. Recently, single-stage WSSS was awakened alleviate problems expensive computational costs complicated training procedures in multi-stage WSSS. However, results such an immature model suffer from background incompleteness object incompleteness. We empirically find that they are caused by insufficiency global context lack local regional contents, respectively. Under these observations, we propose a with only image-level class label supervisions, termed as Weakly Supervised Feature Coupling Network (WS-FCN), which can capture multi-scale formed adjacent feature grids, encode fine-grained spatial information low-level features into high-level ones. Specifically, flexible aggregation module is proposed different granular spaces. Besides, semantically consistent fusion bottom-up parameter-learnable fashion aggregate contents. Based on two modules, WS-FCN lies self-supervised end-to-end fashion. Extensive experimental challenging PASCAL VOC 2012 MS COCO 2014 demonstrate effectiveness efficiency WS-FCN, achieve state-of-the-art 65.02\% 64.22\% mIoU val set test set, 34.12\% The code weight released at:https://github.com/ChunyanWang1/ws-fcn.