LLLMWIKI
正文研究地图阅读入口元数据

来源笔记 · 更新于 2026-09-25

Matryoshka Gaussian Splatting按前缀长度选择高斯预算

阅读原论文

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

一句话结论

已知相机的逐场训练将高斯按不透明度排序，随机选择前缀并同时监督全量渲染，使一个训练完的资产能按预算截取。作者稿p.8 Table1显示m360全量PSNR28.20，比同表3DGS-MCMC的28.40略低；LoD面积指标需按自定义归一化范围理解。

输入、机制与版本

静态场景多视及相机，gsplat的3DGS-MCMC为底座；50k步、容量上限5M，每轮重排图元并随机选择r比例前缀，前缀和全量分别渲染并等权重L1+D-SSIM优化。渲染时截断前k图元，属于一次训练后的连续预算选择，训练仍依赖逐场相机与多次渲染（作者v2 pp.5–8 Eqs.3–9）。出版社正式章DOI 10.1007/978-3-032-36984-0_18；所存17页原件是作者arXiv v2，出版社主PDF请求重定向HTML，官方补充请求曾超时，页码均指所存作者稿。

主表、冲突与消融

作者v2 p.8 Table1 m360九场：本法最高预算28.20 PSNR/.841 SSIM/.130 LPIPS，AUCfps64.81/AUCsplats77.79；3DGS-MCMC全量28.40/.843/.133，CLoD-3DGS为27.44/.814/.215及AUCfps46.38。T&T本法24.56<Octree-GS24.59，Deep Blending28.41<Octree-GS30.35，BungeeNeRF27.13<Octree-GS28.23；整体面积增益不能代替满容量画质。AUC在PSNR[14,32]、SSIM[.35,.92]、LPIPS[.06,.60]等人为区间归一化，定义见p.7 Eqs.9。p.9文字声称m360 CLoD与本法AUCfps为28.94/54.46，与Table1的46.38/64.81冲突，保留待版本核。p.13 Table2 bicycle单场消融：仅前缀24.97 PSNR/.253 LPIPS/AUCfps56.58；前缀+全量25.43/.175/66.87；升序24.91/30.53，对降序25.47/68.62，不能据单场推全场。

成本与边界

50k步的排序、两次渲染、SfM初始化、训练总时长与峰显存应计入部署前成本；训练好资产才具备按前缀运行的能力，FPS/图元预算跨硬件须重测。仓库GitHub API标识Apache-2.0源码许可，数据另核。未进行GPU复现。对照MCMC的逐场底座、EcoSplat的前馈指定预算和GlobalSplat的固定容量前馈预测。独立证据审计见raw/ingest/2026-09-25-matryoshka-gaussian-splatting/analysis.md，跨文判断见三维高斯泼溅：表示、优化、几何与动态任务。

元数据
{
  "id": "2026-09-25-matryoshka-gaussian-splatting",
  "type": "source",
  "title": "Matryoshka Gaussian Splatting：按前缀长度选择高斯预算",
  "status": "reviewed",
  "note_status": "reviewed",
  "created": "2026-09-25",
  "updated": "2026-09-25",
  "source_type": "paper",
  "venue": "ECCV 2026",
  "authors": "[Zhilin Guo, Boqiao Zhang, Hakan Aktas, Kyle Fogarty, Jeffrey Hu, Nursena Koprucu Aslan, Wenzhao Li, Canberk Baykal, Albert Miao, Josef Bengtson, Chenliang Zhou, Weihao Xia, Cristina Nader Vasconcelos, Cengiz Oztireli]",
  "published_at": "2026-09-25",
  "ingested_at": "2026-09-25",
  "canonical_links": [
    "https://link.springer.com/chapter/10.1007/978-3-032-36984-0_18",
    "https://arxiv.org/pdf/2603.19234",
    "https://github.com/ZhilinGuo/matryoshka-gaussian-splatting"
  ],
  "raw_entry": "raw/ingest/2026-09-25-matryoshka-gaussian-splatting/",
  "topics": "[topics/3d-gaussian-splatting]",
  "entities": [],
  "claims": [],
  "questions": [],
  "tags": [
    "near-cvpr-2025",
    "3d-vision",
    "gaussian-splatting",
    "primary-source"
  ]
}