Enhanced online CAM: Single-stage weakly supervised semantic segmentation via collaborative guidance

Bingfeng Zhang, Xuru Gao, Siyue Yu, Weifeng Liu*

*Corresponding author for this work

Research output: Contribution to journalArticlepeer-review

Abstract

Weakly supervised semantic segmentation with image-level annotations usually adopts multi-stage approaches, where high-quality offline CAM is generated as pseudo labels for further training, leading to a complex training process. Instead, current single-stage approaches, directly learning to segment objects with online CAM from image-level supervision, are more elegant. The quality of CAM critically determines the final segmentation performance. However, how to generate high-quality online CAM has not been deeply studied in existing single-stage methods. In this paper, we propose a new single-stage framework to mine more relative target features for enhanced online CAM. Specifically, we design a novel Collaborative Guidance Mechanism, where a prior guidance block uses the original CAM to produce class-specific feature representations, improving the quality of online CAM. However, such a prior is sensitive to discriminative regions of objects. Thus, we further propose a prior fusion block, in which the online segmentation prediction and the original CAM are fused to strengthen the prior guidance. Extensive experiments show that our approach achieves new state-of-the-art performance on both PASCAL VOC 2012 and MS COCO 2014 datasets, outperforming recent single-stage methods by a clear margin. Code is available at https://github.com/1rua11/CGM

Original languageEnglish
Article number110787
JournalPattern Recognition
Volume156
DOIs
Publication statusPublished - Dec 2024
Externally publishedYes

Keywords

  • CAM
  • Semantic segmentation
  • Single-stage
  • Weakly supervised learning

Fingerprint

Dive into the research topics of 'Enhanced online CAM: Single-stage weakly supervised semantic segmentation via collaborative guidance'. Together they form a unique fingerprint.

Cite this