Scholarlink

<p>Non-line-of-sight (NLOS) imaging has emerged as an essential frontier in computational imaging, but traditional active NLOS methods rely on extensive scanning, hindering their use in real-time applications such as robotics and autonomous driving. Using under-scanning measurements (UM) is an effective strategy to accelerate acquisition, but it often leads to degraded reconstruction quality due to insufficient spatial-temporal information. To address this challenge, we propose a dual-model guided under-scanning reconstruction framework that significantly improves both efficiency and fidelity. Specifically, a spatio-temporal recovery module (STRM) first compensates for the missing information in UM by employing 3D pyramid pooling and window attention, transforming them into sufficient-scanning measurements (SM) before reconstruction. Subsequently, the dual-enhancement reconstruction module (DERM) integrates the light-cone transform (LCT) model, which leverages geometrical-optics principles to capture global structures, and the frequency–wavenumber migration (FK) model, which exploits wave-propagation characteristics to recover fine textures simultaneously. Both branches are equipped with refinement modules employing multi-scale convolution and spatial dynamic selection to enhance feature representation, and their refined features are further fused through an adaptive fusion module to fully exploit their complementarity. Extensive experiments demonstrate that our method can effectively reconstruct hidden objects with substantially reduced scanning time, while maintaining efficient inference and reasonable memory usage on modern GPUs. The code is available at <a href="https://github.com/yand90589-rgb/nlos.">https://github.com/yand90589-rgb/nlos.</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

Dual-model guided active NLOS imaging with under-scanning measurements

  • 期刊论文

  • 发表日期：

    2026年2月19日

查看全文
  • Zhihang Yan,
  • Hao Liu,
  • Mengge Liu,
  • Sai Zhang,
  • Huimin Wang,
  • Shaohui Jin,
  • Mingliang Xu

摘要

Non-line-of-sight (NLOS) imaging has emerged as an essential frontier in computational imaging, but traditional active NLOS methods rely on extensive scanning, hindering their use in real-time applications such as robotics and autonomous driving. Using under-scanning measurements (UM) is an effective strategy to accelerate acquisition, but it often leads to degraded reconstruction quality due to insufficient spatial-temporal information. To address this challenge, we propose a dual-model guided under-scanning reconstruction framework that significantly improves both efficiency and fidelity. Specifically, a spatio-temporal recovery module (STRM) first compensates for the missing information in UM by employing 3D pyramid pooling and window attention, transforming them into sufficient-scanning measurements (SM) before reconstruction. Subsequently, the dual-enhancement reconstruction module (DERM) integrates the light-cone transform (LCT) model, which leverages geometrical-optics principles to capture global structures, and the frequency–wavenumber migration (FK) model, which exploits wave-propagation characteristics to recover fine textures simultaneously. Both branches are equipped with refinement modules employing multi-scale convolution and spatial dynamic selection to enhance feature representation, and their refined features are further fused through an adaptive fusion module to fully exploit their complementarity. Extensive experiments demonstrate that our method can effectively reconstruct hidden objects with substantially reduced scanning time, while maintaining efficient inference and reasonable memory usage on modern GPUs. The code is available at https://github.com/yand90589-rgb/nlos.