Scholarlink

<p>RGB scene parsing plays a vital role in applications such as autonomous driving, smart surveillance, and robot vision. Despite recent progress, many existing methods still struggle to balance local feature precision and global context modeling. To address this, we propose SPCFormer, a novel transformer-based architecture that integrates two key components: the Spatial Perception Correction Encoder (SPCE) and the Cross-fusion Attention Decoder (CFAD). SPCE enhances local feature extraction by adaptively correcting spatial offsets via a lightweight, nonlinear modulation function, while CFAD improves global semantic understanding through efficient multi-stage feature fusion. These designs allow SPCFormer to achieve strong performance with a compact model size, making it suitable for dense prediction tasks under resource constraints. Extensive experiments on single-task and multi-task scene parsing demonstrate that SPCFormer achieves state-of-the-art accuracy on PASCAL-Context, NYUDv2, and SUN RGB-D.</p>

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

Spcformer: spatial perception correction transformer for semantic segmentation of scene parsing

  • 期刊论文

  • 发表日期：

    2025年8月21日

查看全文
  • Zhengan Lu,
  • Zhuang Zhou,
  • Shuobin Wei,
  • Zizhao Yuan,
  • Binghua Su

摘要

RGB scene parsing plays a vital role in applications such as autonomous driving, smart surveillance, and robot vision. Despite recent progress, many existing methods still struggle to balance local feature precision and global context modeling. To address this, we propose SPCFormer, a novel transformer-based architecture that integrates two key components: the Spatial Perception Correction Encoder (SPCE) and the Cross-fusion Attention Decoder (CFAD). SPCE enhances local feature extraction by adaptively correcting spatial offsets via a lightweight, nonlinear modulation function, while CFAD improves global semantic understanding through efficient multi-stage feature fusion. These designs allow SPCFormer to achieve strong performance with a compact model size, making it suitable for dense prediction tasks under resource constraints. Extensive experiments on single-task and multi-task scene parsing demonstrate that SPCFormer achieves state-of-the-art accuracy on PASCAL-Context, NYUDv2, and SUN RGB-D.