Scholarlink

Weakly Supervised Semantic Segmentation (WSSS) methods based on image-level labeling alleviate the burden of annotation due to image-level labels only providing object categories. However, the generated Class Activation Maps (CAM) can only localize the most discriminative region rather than the entire object region. To remedy this issue, a framework of WSSS algorithms for Cross-Domain Calibration and Boundary Denoising Network (CD-CBN) is presented. Specifically, a Spatial Feature Calibration Network (SFCN) is proposed to align cross-dimensional features with class prototypes, focusing on intra-class feature consistency. Then, a Class-Specific Distance Model (CSDM) is adopted to separate features from different classes, and feature activation in the object region surpasses the background area. Finally, a Full-domain-aware Noise Reduction Model (FNRM) is designed to refine the object boundary pixels by capturing global contextual features and further filtering out pixel-level noise. A comprehensive experimental evaluation of the highly challenging Pascal VOC 2012 dataset and MS COCO 2014 is presented in this study, illustrating the effectiveness of our suggested approach.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Cross-Domain Calibration and Boundary Denoising Network for Weakly Supervised Semantic Segmentation

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Zhoufeng Liu,
  • Bingrui Li,
  • Shumin Ding,
  • Jiangtao Xi,
  • Chunlei Li

摘要

Weakly Supervised Semantic Segmentation (WSSS) methods based on image-level labeling alleviate the burden of annotation due to image-level labels only providing object categories. However, the generated Class Activation Maps (CAM) can only localize the most discriminative region rather than the entire object region. To remedy this issue, a framework of WSSS algorithms for Cross-Domain Calibration and Boundary Denoising Network (CD-CBN) is presented. Specifically, a Spatial Feature Calibration Network (SFCN) is proposed to align cross-dimensional features with class prototypes, focusing on intra-class feature consistency. Then, a Class-Specific Distance Model (CSDM) is adopted to separate features from different classes, and feature activation in the object region surpasses the background area. Finally, a Full-domain-aware Noise Reduction Model (FNRM) is designed to refine the object boundary pixels by capturing global contextual features and further filtering out pixel-level noise. A comprehensive experimental evaluation of the highly challenging Pascal VOC 2012 dataset and MS COCO 2014 is presented in this study, illustrating the effectiveness of our suggested approach.