Scholarlink

<p>Fisheye lenses, with their ultra-wide field of view, are invaluable in computer vision tasks such as video surveillance, autonomous driving, and virtual reality. However, the severe radial geometric distortion they introduce poses significant challenges. This paper introduces MaFIR, a novel framework for fisheye image rectification built upon orthogonal distortion decoupling and dynamic feature optimization. Specifically, it leverages Manhattan self-attention (MaSA) to introduce explicit spatial priors for orthogonal optical flow regression and employs a Partitioned Group Shuffle Attention (PGSA) module to achieve dynamic feature reweighting, effectively mitigating edge blurring caused by non-uniform stretching. MaFIR employs a two-stage training strategy to decouple geometric distortion from content features, enhancing texture representation and model efficiency. Experimental results on the Places365 dataset demonstrate MaFIR’s superiority over mainstream models, achieving a PSNR of 25.14&#xa0;dB and an SSIM of 0.91, while processing 1024 × 1024 images in just 32.77&#xa0;ms. Code is available at <a href="https://github.com/GovinerL/MaFIR/blob/master">https://github.com/GovinerL/MaFIR/blob/master</a>.</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

MaFIR: high-fidelity fisheye image rectification via Manhattan self-attention and dynamic feature optimization

  • 期刊论文

  • 发表日期：

    2026年5月31日

查看全文
  • Wenzhuo Gao,
  • Bo Zhang,
  • Guoping Wang,
  • Xin Ren,
  • Lei Zhu,
  • Yang Pan,
  • Wenteng Cao

摘要

Fisheye lenses, with their ultra-wide field of view, are invaluable in computer vision tasks such as video surveillance, autonomous driving, and virtual reality. However, the severe radial geometric distortion they introduce poses significant challenges. This paper introduces MaFIR, a novel framework for fisheye image rectification built upon orthogonal distortion decoupling and dynamic feature optimization. Specifically, it leverages Manhattan self-attention (MaSA) to introduce explicit spatial priors for orthogonal optical flow regression and employs a Partitioned Group Shuffle Attention (PGSA) module to achieve dynamic feature reweighting, effectively mitigating edge blurring caused by non-uniform stretching. MaFIR employs a two-stage training strategy to decouple geometric distortion from content features, enhancing texture representation and model efficiency. Experimental results on the Places365 dataset demonstrate MaFIR’s superiority over mainstream models, achieving a PSNR of 25.14 dB and an SSIM of 0.91, while processing 1024 × 1024 images in just 32.77 ms. Code is available at https://github.com/GovinerL/MaFIR/blob/master.