Scholarlink

<p>Despite significant advancements in novel view synthesis achieved by 3D Gaussian splatting, its reliance on dense input views and high storage requirements limit its applicability. To address this challenge, this study presents LFVGS, a lightweight novel view synthesis framework based on 3DGS technology. LFVGS achieves low-cost, high-quality novel view synthesis with the constraint of a limited number of input views (as few as three), while significantly reducing storage requirements. To achieve high performance with sparse inputs, we introduce two key strategies. First, to mitigate the problem of insufficient initial inputs, we leverage distance scores and opacity thresholds to dynamically and densely insert new Gaussians among the initial Gaussian primitives, thereby avoiding the issue of inadequate representation under sparse view constraints. Second, to strengthen geometric constraints imposed by sparse input views, we incorporate depth priors as supervision, utilizing global-to-local depth regularization to reduce discrepancies between the rendered depth and depth priors, thereby correcting scene geometry. Finally, by employing a simple MLP to represent view-dependent color, the synthesis quality of complex reflective regions and local object contours is significantly improved, while model storage overhead is markedly reduced. Experimental results demonstrate that LFVGS achieves excellent overall performance on the LLFF, MipNeRF360, and Shiny datasets, enabling a fully end-to-end, lightweight optimization process. The code is available at: <a href="https://github.com/LeeXiaoTong1/LFVGS">https://github.com/LeeXiaoTong1/LFVGS</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

LFVGS: lightweight Gaussian splatting method for few-shot view synthesis

  • 期刊论文

  • 发表日期：

    2025年3月5日

查看全文
  • Xiaotong Li,
  • Kefeng Li,
  • Guangyuan Zhang,
  • Zhenfang Zhu,
  • Peng Wang,
  • Zhenfei Wang,
  • Chen Fu,
  • Yongshuo Zhang,
  • Zhiming Fan,
  • Yongpeng Zhao

摘要

Despite significant advancements in novel view synthesis achieved by 3D Gaussian splatting, its reliance on dense input views and high storage requirements limit its applicability. To address this challenge, this study presents LFVGS, a lightweight novel view synthesis framework based on 3DGS technology. LFVGS achieves low-cost, high-quality novel view synthesis with the constraint of a limited number of input views (as few as three), while significantly reducing storage requirements. To achieve high performance with sparse inputs, we introduce two key strategies. First, to mitigate the problem of insufficient initial inputs, we leverage distance scores and opacity thresholds to dynamically and densely insert new Gaussians among the initial Gaussian primitives, thereby avoiding the issue of inadequate representation under sparse view constraints. Second, to strengthen geometric constraints imposed by sparse input views, we incorporate depth priors as supervision, utilizing global-to-local depth regularization to reduce discrepancies between the rendered depth and depth priors, thereby correcting scene geometry. Finally, by employing a simple MLP to represent view-dependent color, the synthesis quality of complex reflective regions and local object contours is significantly improved, while model storage overhead is markedly reduced. Experimental results demonstrate that LFVGS achieves excellent overall performance on the LLFF, MipNeRF360, and Shiny datasets, enabling a fully end-to-end, lightweight optimization process. The code is available at: https://github.com/LeeXiaoTong1/LFVGS