Scholarlink

<p>Neural Radiance Fields (NeRF), under dense view coverage, has sparked a new wave in 3D reconstruction with its photo-realistic rendering quality and simple yet efficient network architecture. However, the rendering quality for novel views significantly decreases when the number of input views is drastically reduced. We observe that the degenerate solution of scene reconstruction in sparse views is caused by the misestimation of the geometric structure and detail information. In this paper, we present a novel framework, LG-NeRF, that employs local and global rays to jointly optimize the performance of NeRF under sparse view settings. First, we propose a matching-based local ray optimization scheme, which extracts the scene’s geometric consistency from rough depth maps, providing a reliable geometric structure for the NeRF training process. Second, a global ray optimization strategy based on self-generated resources is employed by retrieving high-frequency information from invisible viewpoints, which allows NeRF to reconstruct more accurate details. Experiments demonstrate that our method significantly improves the rendering quality of novel views under sparse input conditions on LLFF and DTU datasets, achieving state-of-the-art results.</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

LG-NeRF: Local-global ray joint optimization for few-shot novel view synthesis

  • 期刊论文

  • 发表日期：

    2025年9月11日

查看全文
  • Jian Ge,
  • Liang Yuan,
  • Shumei Bao,
  • Wendong Xiao,
  • Teng Ran,
  • Kai Lv,
  • Jianping Cui

摘要

Neural Radiance Fields (NeRF), under dense view coverage, has sparked a new wave in 3D reconstruction with its photo-realistic rendering quality and simple yet efficient network architecture. However, the rendering quality for novel views significantly decreases when the number of input views is drastically reduced. We observe that the degenerate solution of scene reconstruction in sparse views is caused by the misestimation of the geometric structure and detail information. In this paper, we present a novel framework, LG-NeRF, that employs local and global rays to jointly optimize the performance of NeRF under sparse view settings. First, we propose a matching-based local ray optimization scheme, which extracts the scene’s geometric consistency from rough depth maps, providing a reliable geometric structure for the NeRF training process. Second, a global ray optimization strategy based on self-generated resources is employed by retrieving high-frequency information from invisible viewpoints, which allows NeRF to reconstruct more accurate details. Experiments demonstrate that our method significantly improves the rendering quality of novel views under sparse input conditions on LLFF and DTU datasets, achieving state-of-the-art results.