Scholarlink

<p>Tibetan temple cultural symbols, rich in cultural significance, are facing dual threats from the harsh plateau climate and human activity, making it urgent to apply digital technologies for their preservation and transmission. Challenges such as texture blurring, geometric incompleteness, and poor adaptability to dynamic lighting can be observed in the 3D reconstruction of Tibetan temple cultural symbols. To solve these problems, we propose a geometry-enhanced and direction-aware rendering framework built upon 3D Gaussian splatting. The model integrates multi-scale 3D low-pass filtering and 2D Mip filtering to improve the fidelity of Gaussian representation, and employs a direction-aware MLP to dynamically modulate Gaussian attributes, enhancing rendering consistency under complex illumination. Additionally, we design a depth-based geometry optimizer to refine uncertain structural regions by generating accurate Gaussian distributions. Experiments on public datasets (NeRF360, Tanks &amp; Temples, DeepBlending) and our custom Tibetan temple symbol dataset validate the effectiveness of our model. Our model outperforms 3DGS on Tanks &amp; Temples, DeepBlending datasets with a PSNR gain of 0.90&#xa0;dB, while reducing LPIPS by 16.7% demonstrates better geometric fidelity in slender edge reconstruction. If the computational power of supercomputing platforms can be leveraged, this research will significantly accelerate dataset processing and rendering, enabling real-time rendering of complex textures and more efficient optimization, thereby providing a high-precision 3D reconstruction solution for the preservation of Tibetan temple cultural symbols.</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

DirecGeo-GS: direction-aware and geometry-enhanced Gaussian splatting for Tibetan temple symbols

  • 期刊论文

  • 发表日期：

    2025年11月12日

查看全文
  • Wang Yitong,
  • Yang Xiaobo,
  • Zhou Mingqiang,
  • Qiu Kexin,
  • Wang Jiashuo

摘要

Tibetan temple cultural symbols, rich in cultural significance, are facing dual threats from the harsh plateau climate and human activity, making it urgent to apply digital technologies for their preservation and transmission. Challenges such as texture blurring, geometric incompleteness, and poor adaptability to dynamic lighting can be observed in the 3D reconstruction of Tibetan temple cultural symbols. To solve these problems, we propose a geometry-enhanced and direction-aware rendering framework built upon 3D Gaussian splatting. The model integrates multi-scale 3D low-pass filtering and 2D Mip filtering to improve the fidelity of Gaussian representation, and employs a direction-aware MLP to dynamically modulate Gaussian attributes, enhancing rendering consistency under complex illumination. Additionally, we design a depth-based geometry optimizer to refine uncertain structural regions by generating accurate Gaussian distributions. Experiments on public datasets (NeRF360, Tanks & Temples, DeepBlending) and our custom Tibetan temple symbol dataset validate the effectiveness of our model. Our model outperforms 3DGS on Tanks & Temples, DeepBlending datasets with a PSNR gain of 0.90 dB, while reducing LPIPS by 16.7% demonstrates better geometric fidelity in slender edge reconstruction. If the computational power of supercomputing platforms can be leveraged, this research will significantly accelerate dataset processing and rendering, enabling real-time rendering of complex textures and more efficient optimization, thereby providing a high-precision 3D reconstruction solution for the preservation of Tibetan temple cultural symbols.