Scholarlink

<p>This study systematically benchmarks eight real-time semantic segmentation models—ENet, ICNet, BiSeNet V1, Fast-SCNN, BiSeNet V2, PIDNet, DDRNet, and TopFormer, across Stanford Background, CamVid, Cityscapes, and Indian Driving Dataset (IDD) under a consistent protocol. Performance is measured using mean IoU, inference speed (FPS), and computational efficiency (FLOPs and parameter count). Results reveal trade-offs driven by dataset characteristics and architectural design. Multi-branch networks with distinct high-resolution spatial and global contextual pathways, such as PIDNet (83.19% mIoU on CamVid) and BiSeNet V1 (64.78% mIoU on IDD Lite), generalize better to unstructured environments. Ultra-lightweight models like Fast-SCNN and ICNet achieve higher inference speeds (Fast-SCNN reaching a peak inference speed of 1279 FPS on an NVIDIA RTX 4070) at the cost of segmentation quality. These findings offer practical guidance for real-time model selection.</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

Benchmarking real-time semantic segmentation models for traffic scene understanding across structured and unstructured environments

  • 期刊论文

  • 发表日期：

    2026年7月14日

查看全文
  • Mostafizur Rahman,
  • Nairiti Dutta,
  • Snehil Jha,
  • Alexy Bhowmick,
  • Karsten Schmidt

摘要

This study systematically benchmarks eight real-time semantic segmentation models—ENet, ICNet, BiSeNet V1, Fast-SCNN, BiSeNet V2, PIDNet, DDRNet, and TopFormer, across Stanford Background, CamVid, Cityscapes, and Indian Driving Dataset (IDD) under a consistent protocol. Performance is measured using mean IoU, inference speed (FPS), and computational efficiency (FLOPs and parameter count). Results reveal trade-offs driven by dataset characteristics and architectural design. Multi-branch networks with distinct high-resolution spatial and global contextual pathways, such as PIDNet (83.19% mIoU on CamVid) and BiSeNet V1 (64.78% mIoU on IDD Lite), generalize better to unstructured environments. Ultra-lightweight models like Fast-SCNN and ICNet achieve higher inference speeds (Fast-SCNN reaching a peak inference speed of 1279 FPS on an NVIDIA RTX 4070) at the cost of segmentation quality. These findings offer practical guidance for real-time model selection.