LLLMWIKI
ArticleResearch mapReading portalMetadata

Source note · Updated 2026-09-25

FastGS多视误差驱动的增密、剪枝与训练加速

Read the paper

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

English translation pending — this page shows the Chinese original.

一句话结论

CVPR 2026正式版在已标定逐场训练中，以多视图像素误差筛选增密/剪枝并收紧tile包围盒；RTX4090、30K步下训练时间大幅缩短，但基础版在部分感知指标上退化，计时口径不含完整采集/SfM成本（主文Tables1/7，补充Table12）。

输入与机制

输入多视角RGB、COLMAP/SfM相机与稀疏点，输出优化的高斯及新视角RGB。VCD在随机K=10训练视图中对归一化L1高误差像素与高斯足迹求重叠，并结合原梯度门控增密；VCP把高误差足迹与全图L1/SSIM乘积用于筛剪。CB在Speedy-Splat snugbox基础上缩小高斯—tile交集。基底含Taming per-splat并行反传/SH调度和AbsGS绝对梯度；这些实现模块也贡献速度（主文§4 pp.26097–26098；补充§8–9 pp.1–3）。

结果、消融与边界

  • 主文p.26099 Table1：Mip-NeRF360 FastGS基础版1.93min/PSNR27.56/SSIM.797/LPIPS.261/0.40M/579FPS；DashGaussian6.35min/27.73/.817/.218/2.40M/155FPS。Big版3.58min/27.93/.820/.216；“相当质量”是取舍主张，基础版LPIPS较差。Deep Blending比原3DGS 1.28 vs19.77min（该表15.45倍），RTX4090同表重跑、逐场30K步。
  • p.26101 Table7基于3DGS-accel 7.10min/27.46/.226 LPIPS、单VCD3.53min/27.69/.259、单VCP5.32min/27.70/.228、全合1.93min/27.56/.261。补充p.4 Table14 τd=5→100使T&T PSNR24.15→23.26；阈值有质量代价。补充Table12 m360峰值显存5.21GB/存储99MB为作者测得单场统计，仓库README建议24GB运行paper-quality。
  • 跨任务Table4 T&T PGSR+FastGS训练32.28→15.74min，但几何F1 .57→.55、LPIPS .149→.190；补充Table9 Replica Photo-SLAM 5.03→1.86min、LPIPS .026→.042，没有ATE/RPE。主文§5.5承认稠密前馈输出的短时后训练困难。误差覆盖分数是启发式代理，需留一图元测试真实边际贡献；场景级多种子、相机/SfM和同质量速度曲线未见。

原始链接与证据

CVF正式主文10页、官方ZIP抽出的9页补充PDF、代码入口见frontmatter与raw/ingest/2026-09-25-fastgs/links.yaml；详证及SHA256见raw/ingest/2026-09-25-fastgs/analysis.md。仓库顶层MIT，另有原3DGS研究用途许可及第三方代码；独立核许可证。无本地GPU复现。

相关页面

3D Gaussian Splatting: Representation, Optimization, Geometry, and Dynamic Scenes；3D Gaussian Splatting：从稀疏点到实时新视角渲染、3DGS as MCMC：用高斯重定位改写密度控制、Mip-Splatting：跨尺度采样与三维高斯抗混叠。训练速度、质量、几何与SLAM定位各按不同任务口径验收。

Metadata
{
  "id": "2026-09-25-fastgs",
  "type": "source",
  "title": "FastGS：多视误差驱动的增密、剪枝与训练加速",
  "status": "reviewed",
  "note_status": "reviewed",
  "created": "2026-09-25",
  "updated": "2026-09-25",
  "source_type": "paper",
  "venue": "CVPR 2026",
  "authors": "[Shiwei Ren, Tianci Wen, Yongchun Fang, Biao Lu]",
  "published_at": "2026-06-01",
  "ingested_at": "2026-09-25",
  "canonical_links": [
    "https://openaccess.thecvf.com/content/CVPR2026/html/Ren_FastGS_Training_3D_Gaussian_Splatting_in_100_Seconds_CVPR_2026_paper.html",
    "https://openaccess.thecvf.com/content/CVPR2026/papers/Ren_FastGS_Training_3D_Gaussian_Splatting_in_100_Seconds_CVPR_2026_paper.pdf",
    "https://openaccess.thecvf.com/content/CVPR2026/supplemental/Ren_FastGS_Training_3D_CVPR_2026_supplemental.zip",
    "https://github.com/fastgs/FastGS"
  ],
  "raw_entry": "raw/ingest/2026-09-25-fastgs/",
  "topics": "[topics/3d-gaussian-splatting]",
  "entities": [],
  "claims": [],
  "questions": [],
  "tags": [
    "near-cvpr-2025",
    "3d-vision",
    "gaussian-splatting",
    "efficiency",
    "primary-source"
  ]
}