Scholarlink

<p>3D tooth reconstruction is crucial for digital dentistry. Intraoral endoscopic images often contain specular highlights and uneven sampling, which make vanilla 3D Gaussian Splatting (3DGS) prone to floating artifacts and noise, thereby blurring boundaries. We propose a near real-time method tailored to intraoral endoscopy. It iteratively optimizes a large set of Gaussian primitives in a 3DGS pipeline and renders them via GPU-parallel differentiable rasterization. First, we introduce a semantic mask-guided pruning scheme. It uses Mask Coverage Ratio (MCR) with adaptive density pruning inside the mask to preserve tooth structures, and hybrid pruning outside the mask to remove redundant Gaussians. With 30k iterations, we improve PSNR by 8.97% and SSIM by 2.32%, and reduce LPIPS by 11.08% on average across three datasets.</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

Semantic mask-guided adaptive density pruning and local multi-scale regularization for intraoral 3D tooth reconstruction

  • 期刊论文

  • 发表日期：

    2026年2月13日

查看全文
  • Hong-an Li,
  • Chenjia Zhu

摘要

3D tooth reconstruction is crucial for digital dentistry. Intraoral endoscopic images often contain specular highlights and uneven sampling, which make vanilla 3D Gaussian Splatting (3DGS) prone to floating artifacts and noise, thereby blurring boundaries. We propose a near real-time method tailored to intraoral endoscopy. It iteratively optimizes a large set of Gaussian primitives in a 3DGS pipeline and renders them via GPU-parallel differentiable rasterization. First, we introduce a semantic mask-guided pruning scheme. It uses Mask Coverage Ratio (MCR) with adaptive density pruning inside the mask to preserve tooth structures, and hybrid pruning outside the mask to remove redundant Gaussians. With 30k iterations, we improve PSNR by 8.97% and SSIM by 2.32%, and reduce LPIPS by 11.08% on average across three datasets.