LLLMWIKI
ArticleResearch mapReading portalMetadata

Source note · Updated 2026-09-25

CoMe用置信度和方差约束从高斯提取网格

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

逐场有位姿多图重建中，CoMe把每高斯自监督置信度用于光度损失及密度控制，再约束逐高斯颜色/法线，减少假几何。作者同一重跑协议Tanks & Temples六场F1 .521、RTX4090优化18min，对SOF .474/17min；抽网格另约6.6min，输入密度与场景评估边界仍关键（作者稿p.10 Table1、附录p.27）。

机制、实验与版本

在SOF 3D高斯底座上，alpha混合置信度抑制难解释区域光度梯度而保留几何损失；低置信度处提高高斯分裂阈值，亮度项使用单图appearance修正，D-SSIM结构/对比仍用原RGB。沿光线逐高斯颜色/法线方差项约束半透明前后重影，Marching Tetrahedra抽不受界限网格（作者稿pp.5–9 Eqs.8–15）。正式Springer DOI 10.1007/978-3-032-37435-6_17；保存34页arXiv v2作者完整稿和17页官方补充，本文页码对应作者稿。

结果、成本与反证

ScanNet++六选场F1 .668 vs SOF .615、PGSR .631；T&T消融从SOF .474→appearance .493→confidence .509→color-var .519→normal-var .521（作者稿pp.10–11 Tables2–3）。官方作者附录p.20 Table5去置信度增密调整，图元1.27M→1.53M、F1 .521→.516。DTU十五场Chamfer .65高于PGSR .55，先验依赖GeoSVR同T&T原评估F1 .546高于本法.521（作者稿附录pp.26、30 Tables6–7）；MIlo在Mip-NeRF360的mesh-NVS PSNR23.869也高于本法23.639（附录p.34 Table9）。Barn只换预处理本法F1 .534→.616，故论文间GT、体素和输入拆分决定可比性（附录p.21）。主实验均按作者4090优化计时，不含SfM、部分后处理；高峰显存未报。严重遮挡和远景缺细节，密采依赖见p.14；代码许可待核。独立逐页分析raw/ingest/2026-09-25-come/analysis.md。与2DGS、MetroGS按训练卡、几何输出和尺度分账；总览见3D Gaussian Splatting: Representation, Optimization, Geometry, and Dynamic Scenes。

Metadata
{
  "id": "2026-09-25-come",
  "type": "source",
  "title": "CoMe：用置信度和方差约束从高斯提取网格",
  "status": "reviewed",
  "note_status": "reviewed",
  "created": "2026-09-25",
  "updated": "2026-09-25",
  "source_type": "paper",
  "venue": "ECCV 2026",
  "authors": "[Lukas Radl, Felix Windisch, Andreas Kurz, Thomas Köhler, Michael Steiner, Markus Steinberger]",
  "published_at": "2026-09-12",
  "ingested_at": "2026-09-25",
  "canonical_links": [
    "https://link.springer.com/chapter/10.1007/978-3-032-37435-6_17",
    "https://arxiv.org/pdf/2603.24725",
    "https://media.springernature.com/original/springer-static/esm/chp%3A10.1007%2F978-3-032-37435-6_17/MediaObjects/700390_1_En_17_MOESM1_ESM.pdf",
    "https://r4dl.github.io/CoMe/"
  ],
  "raw_entry": "raw/ingest/2026-09-25-come/",
  "topics": "[topics/3d-gaussian-splatting]",
  "entities": [],
  "claims": [],
  "questions": [],
  "tags": [
    "near-cvpr-2025",
    "3d-vision",
    "gaussian-splatting",
    "primary-source"
  ]
}