Scholarlink

<p>While 3D Gaussian Splatting (3DGS) has revolutionized real-time photorealistic rendering, extracting high-fidelity geometry from complex specular and semi-transparent scenes remains a significant challenge. We present UniGS, a unified geometry reconstruction framework designed to overcome these limitations. First, we establish a consistent analytic formulation for depth and normal estimation, eliminating biases found in heuristic approximations. Second, to resolve the shape-radiance ambiguity in specular highlights, we propose a Surface Integrity Constraint that prevents geometric collapse by minimizing ray-accumulated depth. Third, to tackle the multi-surface ambiguity in semi-transparent objects, we develop an Anisotropic Opacity strategy coupled with Multi-Layer Depth Extraction to explicitly recover occluded back surfaces. Extensive experiments on five benchmarks, including DTU, BlendedMVS, and other three specular/transparent datasets, demonstrate that UniGS outperforms state-of-the-art methods in geometric accuracy while maintaining real-time efficiency. Our source code and data are available at: <a href="https://anonymous.4open.science/r/UniGS-D77F">https://anonymous.4open.science/r/UniGS-D77F</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

UniGS: unified geometry reconstruction for specular and semi-transparent surfaces via 3D Gaussian splatting

  • 期刊论文

  • 发表日期：

    2026年3月22日

查看全文
  • Nan Min,
  • Zhiyuan Xu,
  • Yuhang Guo,
  • Wei Jiang,
  • Mofei Song

摘要

While 3D Gaussian Splatting (3DGS) has revolutionized real-time photorealistic rendering, extracting high-fidelity geometry from complex specular and semi-transparent scenes remains a significant challenge. We present UniGS, a unified geometry reconstruction framework designed to overcome these limitations. First, we establish a consistent analytic formulation for depth and normal estimation, eliminating biases found in heuristic approximations. Second, to resolve the shape-radiance ambiguity in specular highlights, we propose a Surface Integrity Constraint that prevents geometric collapse by minimizing ray-accumulated depth. Third, to tackle the multi-surface ambiguity in semi-transparent objects, we develop an Anisotropic Opacity strategy coupled with Multi-Layer Depth Extraction to explicitly recover occluded back surfaces. Extensive experiments on five benchmarks, including DTU, BlendedMVS, and other three specular/transparent datasets, demonstrate that UniGS outperforms state-of-the-art methods in geometric accuracy while maintaining real-time efficiency. Our source code and data are available at: https://anonymous.4open.science/r/UniGS-D77F.