LLLMWIKI
ArticleResearch mapReading portalMetadata

Source note · Updated 2026-09-25

MetroGS多卡分布式二维高斯的大场景几何重建

Read the paper

#near-cvpr-2025#3d-vision#gaussian-splatting#primary-source

English translation pending — this page shows the Chinese original.

一句话结论

MetroGS借SfM/预训练pointmap稠密初始化、单目深度到PatchMatch多视的分阶段几何修正、深度引导的Tri-Mip外观和4×3090分布式训练，在城市点云F1与RGB上同表优于CityGSV2；城市航拍F1.677 vs .556，而同表CityGS-X的PSNR27.58高于本法27.52（CVF主文p.6 Table2）。

机制与数据

输入城市已知相机图像、SfM点/匹配及预训练pointmap/单目深度；高斯按设备重分配，弱纹理大覆盖区域定向分裂，后期多尺度PatchMatch深度按重投影过滤并以局部拟合的单目深度回填；Tri-Mip 3D空间特征与每图appearance embedding改善光照差异（主文pp.3–6 Fig.2/Eqs.3–13）。通过中位深度加TSDF导出网格；GauU-Scene、MatrixCity的像素和TSDF voxel/截断按场景变化（主文p.7及补充p.1）。CVF正式CVPR2026主文11页、补充4页，页码按PDF；正式页pp.395–405。

结果、消融与成本

GauU-Scene Russian Building PSNR24.94/F1.585 vs CityGSV2 24.12/.544；Residence24.51/.494 vs23.57/.467；Modern Building27.07/.524 vs25.84/.492（主文p.6 Table1）。Russian Building p.8 Table3从Base .523/134min 到Base+Para .532/68min再到Full .585/106min；去几何F1 .564、去外观.562、去pointmap初始化.577，故完整模型106min不能称68min。补充p.1 Table1在同4卡配置本法106min vs CityGSV2 363min；MatrixCity街景CityGS-X OOM，内存预算影响比较（主文p.6 Table2）。补充p.3 §D指出显存瓶颈和2DGS渲染上限；SfM、pointmap、TSDF的完整链耗时/峰显存未统一报，许可待核。published_at用2026-06-01作为CVPR出版月份占位，CVF页面未标日级日期。本轮未复现GPU，逐页审计raw/ingest/2026-09-25-metrogs/analysis.md。与CoMe的单卡无界网格优化和2DGS的物体尺度基线分账；总览3D Gaussian Splatting: Representation, Optimization, Geometry, and Dynamic Scenes。

Metadata
{
  "id": "2026-09-25-metrogs",
  "type": "source",
  "title": "MetroGS：多卡分布式二维高斯的大场景几何重建",
  "status": "reviewed",
  "note_status": "reviewed",
  "created": "2026-09-25",
  "updated": "2026-09-25",
  "source_type": "paper",
  "venue": "CVPR 2026",
  "authors": "[Kehua Chen, Tianlu Mao, Xinzhu Ma, Hao Jiang, Zehao Li, Zihan Liu, Shuqi Gao, Honglong Zhao, Feng Dai, Yucheng Zhang, Zhaoqi Wang]",
  "published_at": "2026-06-01",
  "ingested_at": "2026-09-25",
  "canonical_links": [
    "https://openaccess.thecvf.com/content/CVPR2026/html/Chen_MetroGS_Efficient_and_Stable_Reconstruction_of_Geometrically_Accurate_High-Fidelity_Large-Scale_CVPR_2026_paper.html",
    "https://openaccess.thecvf.com/content/CVPR2026/papers/Chen_MetroGS_Efficient_and_Stable_Reconstruction_of_Geometrically_Accurate_High-Fidelity_Large-Scale_CVPR_2026_paper.pdf",
    "https://openaccess.thecvf.com/content/CVPR2026/supplemental/Chen_MetroGS_Efficient_and_CVPR_2026_supplemental.pdf",
    "https://m3phist0.github.io/MetroGS/"
  ],
  "raw_entry": "raw/ingest/2026-09-25-metrogs/",
  "topics": "[topics/3d-gaussian-splatting]",
  "entities": [],
  "claims": [],
  "questions": [],
  "tags": [
    "near-cvpr-2025",
    "3d-vision",
    "gaussian-splatting",
    "primary-source"
  ]
}