Scholarlink

<p>Semantic segmentation is essential for safety-critical applications such as autonomous driving and urban robotics, yet achieving high accuracy under strict constraints on computational resources and inference latency remains challenging. Many existing approaches rely on fixed attention patterns and coarse multi-scale fusion, which weakens long-range reasoning and blurs object boundaries at high resolution. We introduce an efficient framework that couples progressive cross-layer attention with complexity-adaptive sparsity (PCAS), a dual-norm convolutional feed-forward network (DN-ConvFFN), and a semantics-guided cross-scale fusion decoder (SGCF). PCAS builds bidirectional dependencies across hierarchical features while allocating computation to a content-adaptive set of informative tokens, substantially reducing floating point operations (FLOPs) without sacrificing accuracy. DN-ConvFFN fuses batch- and layer-normalized paths with a lightweight gate to stabilize optimization and improve parameter efficiency, and SGCF preserves fine detail by injecting high-level semantics to guide spatial selection during multi-scale fusion. Across standard benchmarks, the proposed model delivers a strong trade-off between accuracy and efficiency: on Cityscapes it achieves 80.5% mean intersection over union (mIoU) at 56.4 frames per second (FPS), and on ADE20K it attains 44.3% mIoU with modest compute, meeting the demands of real-time segmentation. Compared to competitive real-time baselines, our approach improves accuracy while cutting FLOPs and parameters by 30.8% and 26.2%, respectively. Comprehensive ablations further verify the contribution of each component and the effectiveness of complexity-adaptive sparsity and progressive attention under efficient constraints.</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

PCASeg: progressive cross-layer attention with adaptive sparsity for real-time semantic segmentation

  • 期刊论文

  • 发表日期：

    2026年4月5日

查看全文
  • Dongpu Li,
  • Xiaosu Xu

摘要

Semantic segmentation is essential for safety-critical applications such as autonomous driving and urban robotics, yet achieving high accuracy under strict constraints on computational resources and inference latency remains challenging. Many existing approaches rely on fixed attention patterns and coarse multi-scale fusion, which weakens long-range reasoning and blurs object boundaries at high resolution. We introduce an efficient framework that couples progressive cross-layer attention with complexity-adaptive sparsity (PCAS), a dual-norm convolutional feed-forward network (DN-ConvFFN), and a semantics-guided cross-scale fusion decoder (SGCF). PCAS builds bidirectional dependencies across hierarchical features while allocating computation to a content-adaptive set of informative tokens, substantially reducing floating point operations (FLOPs) without sacrificing accuracy. DN-ConvFFN fuses batch- and layer-normalized paths with a lightweight gate to stabilize optimization and improve parameter efficiency, and SGCF preserves fine detail by injecting high-level semantics to guide spatial selection during multi-scale fusion. Across standard benchmarks, the proposed model delivers a strong trade-off between accuracy and efficiency: on Cityscapes it achieves 80.5% mean intersection over union (mIoU) at 56.4 frames per second (FPS), and on ADE20K it attains 44.3% mIoU with modest compute, meeting the demands of real-time segmentation. Compared to competitive real-time baselines, our approach improves accuracy while cutting FLOPs and parameters by 30.8% and 26.2%, respectively. Comprehensive ablations further verify the contribution of each component and the effectiveness of complexity-adaptive sparsity and progressive attention under efficient constraints.