Scholarlink

<p>Novel view synthesis from sparse image collections remains a challenging task in computer vision, in scenes with complex spatial structures or sparse viewpoint overlaps. While neural radiance field (NeRF) technology has shown promising results, existing methods often suffer from color ghosting, structural deformations, and holes in rendered images. This paper introduces the Random Patch Radiance Field (RPRF), an extension of TensoRF, which addresses these issues by randomly sampling image patches and computing patch-level structural similarity between rendered and real images. We employ a similarity loss to guide the model in focusing on local regions and introduce variable positional encoding to improve the representation of complex spatial structures. Experimental results on multiple public datasets demonstrate that RPRF outperforms state-of-the-art methods, achieving a 15.4% improvement in PSNR and a 19.6% improvement in SSIM on the complex scene dataset, Tanks and Temples Advanced Scenes. In addition, it also performs well under sparse viewpoints. Our work highlights the potential of random sampling and dynamic encoding in enhancing the robustness and efficiency of radiance field methods for novel view synthesis. Source code is available at <a href="https://github.com/YGH189/RPRF.git">https://github.com/YGH189/RPRF.git</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

Enhancing novel view synthesis with random patch radiance fields

  • 期刊论文

  • 发表日期：

    2026年2月26日

查看全文
  • Weijie Ren,
  • Guohui Yang,
  • Kang Zou,
  • Zhiyu Lyu

摘要

Novel view synthesis from sparse image collections remains a challenging task in computer vision, in scenes with complex spatial structures or sparse viewpoint overlaps. While neural radiance field (NeRF) technology has shown promising results, existing methods often suffer from color ghosting, structural deformations, and holes in rendered images. This paper introduces the Random Patch Radiance Field (RPRF), an extension of TensoRF, which addresses these issues by randomly sampling image patches and computing patch-level structural similarity between rendered and real images. We employ a similarity loss to guide the model in focusing on local regions and introduce variable positional encoding to improve the representation of complex spatial structures. Experimental results on multiple public datasets demonstrate that RPRF outperforms state-of-the-art methods, achieving a 15.4% improvement in PSNR and a 19.6% improvement in SSIM on the complex scene dataset, Tanks and Temples Advanced Scenes. In addition, it also performs well under sparse viewpoints. Our work highlights the potential of random sampling and dynamic encoding in enhancing the robustness and efficiency of radiance field methods for novel view synthesis. Source code is available at https://github.com/YGH189/RPRF.git.