Scholarlink

<p>While 3D Gaussian Splatting (3DGS) has revolutionized novel-view synthesis, accurately recovering reflective surfaces remains a significant challenge due to inherent depth estimation errors and the limited capacity of spherical harmonics in representing high-frequency reflections. In this paper, we propose PGSR-DR, a reflection-aware framework that integrates planar-based Gaussian reconstruction with deferred rendering for high-fidelity geometry and appearance recovery. We first establish a reliable geometric foundation by introducing a depth-calculation method for planar-based Gaussians. Our method eliminates conventional estimation artifacts and incorporates joint depth-normal consistency and multi-view supervision to ensure global structural coherence. To capture intricate specularities, we incorporate a learnable environment map within a deferred rendering pipeline that uses Nvdiffrast for efficient sampling and explicit modeling of view-dependent appearances. Experimental results demonstrate that our method achieves competitive rendering quality and notably improved geometric accuracy for reflective surfaces, with planar-based Gaussian primitives closely adhering to the underlying surfaces while maintaining real-time performance.</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

PGSR-DR: high-fidelity reflective surface reconstruction with planar-based Gaussians and deferred rendering

  • 期刊论文

  • 发表日期：

    2026年7月12日

查看全文
  • Jingfeng Li,
  • Xiaokun Wang,
  • Haokai Zeng,
  • Xingyu Ye,
  • Jiří Kosinka,
  • Alexandru C. Telea,
  • Yalan Zhang,
  • Yanrui Xu

摘要

While 3D Gaussian Splatting (3DGS) has revolutionized novel-view synthesis, accurately recovering reflective surfaces remains a significant challenge due to inherent depth estimation errors and the limited capacity of spherical harmonics in representing high-frequency reflections. In this paper, we propose PGSR-DR, a reflection-aware framework that integrates planar-based Gaussian reconstruction with deferred rendering for high-fidelity geometry and appearance recovery. We first establish a reliable geometric foundation by introducing a depth-calculation method for planar-based Gaussians. Our method eliminates conventional estimation artifacts and incorporates joint depth-normal consistency and multi-view supervision to ensure global structural coherence. To capture intricate specularities, we incorporate a learnable environment map within a deferred rendering pipeline that uses Nvdiffrast for efficient sampling and explicit modeling of view-dependent appearances. Experimental results demonstrate that our method achieves competitive rendering quality and notably improved geometric accuracy for reflective surfaces, with planar-based Gaussian primitives closely adhering to the underlying surfaces while maintaining real-time performance.