LLLMWIKI
ArticleResearch mapReading portalMetadata

Source note · Updated 2026-09-25

SpeeDe3DGS动态高斯时间剪枝与运动分组

Read the paper

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

English translation pending — this page shows the Chinese original.

一句话结论

CVPR 2026在动态DeformableGS上以时间敏感度剪枝、抖动采样与共享SE(3)运动降低高斯和逐帧变形开销。NeRF-DS七场全合FPS54.37→505.60、训练1523.83→625.48秒，PSNR23.80→23.66、LPIPS.1781→.1901；MonoDyGauBench50场训练表基线与作者方法使用不同型号GPU（主文pp.26089–26090 Tables1–2）。

输入与机制

输入带时间的多视/单目视频及相机，先有逐场DeformableGS变形网络/高斯；输出时变图元和RGB。TSP以跨训练时间/位姿的图元投影梯度平方近似敏感度剪掉低贡献图元。TSS在敏感度估计时以β=.1、高斯时间抖动于20K步退火探索邻近时刻，弱化未观测时刻的浮点。GroupFlow以2048控制高斯从首帧轨迹聚类，Umeyama拟合组内SE(3)，把N次变形推断变为J组共享运动（§4 pp.26086–26088 Eq.5–12）。这种局部刚体约束对强非刚体有画质代价，TSS无GT时间监督。

可核结果、反证与成本

  • 主文p.26089 Table1 NeRF-DS七真实场三次平均，基线23.80 PSNR/.8503 SSIM/.1781 LPIPS、54.37FPS、33.21MB/132.22K图元；TSP+TSS 23.81/.8515/.1853、345.24FPS、4.55MB/11.06K；再加GroupFlow全合23.66/.8487/.1901、505.60FPS、21.40MB/11.10K。GPU渲染3090、NeRF-DS训练A5000；GroupFlow额外参数增加剪枝后存储。
  • 主文p.26090 Table2 MonoDyGauBench50场：DeformableGS 24.07 PSNR/.283 LPIPS/20.20FPS/6227.43s，+剪枝23.86/.295/137.01FPS/2850.60s*，+分组23.52/.313/276.91FPS/2461.14s*。4DGS基线23.55/.277、+剪枝22.44/.334、全合21.00/.380，质量损失明显。表*训练用A5000，基线3090，训练比值混入硬件；FPS统一3090。表2中部分速度倍率与原始秒数疑似不符，按原值及硬件解释。
  • 官方补充PDF p.14 Table8 D-NeRF八合成场：基线38.92 PSNR/.0143 LPIPS，全合35.07/.0365；FPS127.47→524.19。补充p.14 Table9 HyperNeRF独立八场仅有速度/大小，基线A6000、剪枝版A5000训练，原数据缺标准NVS质量划分；不据此宣称动态画质持平。p.14 Table10中不剪枝的J2048模型51MB超过基线33.21MB，强剪枝时J2048为21.40MB。

原始链接与边界

CVF正式主文11页、官方补充17页（前11页重印主文、后6页附录）及逐页分析、SHA256见raw/ingest/2026-09-25-speede3dgs/analysis.md。仓库API顶层MIT，依赖与数据授权独立审核；没有本地GPU重跑。需同卡同场景固定seed、划分、J和剪枝阈值，计训练/分组拟合、峰显存、FPS、存储、LPIPS及未见时间定量抖动；视频帧不是独立场景。

相关页面

3D Gaussian Splatting: Representation, Optimization, Geometry, and Dynamic Scenes；Faster-GS：训练实现拆账与数值稳定性。动态任务与静态NVS分支分表，不合并单一排行榜。

Metadata
{
  "id": "2026-09-25-speede3dgs",
  "type": "source",
  "title": "SpeeDe3DGS：动态高斯时间剪枝与运动分组",
  "status": "reviewed",
  "note_status": "reviewed",
  "created": "2026-09-25",
  "updated": "2026-09-25",
  "source_type": "paper",
  "venue": "CVPR 2026",
  "authors": "[Allen Tu, Haiyang Ying, Alex Hanson, Yonghan Lee, Tom Goldstein, Matthias Zwicker]",
  "published_at": "2026-06-01",
  "ingested_at": "2026-09-25",
  "canonical_links": [
    "https://openaccess.thecvf.com/content/CVPR2026/html/Tu_SpeeDe3DGS_Speedy_Deformable_3D_Gaussian_Splatting_with_Temporal_Pruning_and_CVPR_2026_paper.html",
    "https://openaccess.thecvf.com/content/CVPR2026/papers/Tu_SpeeDe3DGS_Speedy_Deformable_3D_Gaussian_Splatting_with_Temporal_Pruning_and_CVPR_2026_paper.pdf",
    "https://openaccess.thecvf.com/content/CVPR2026/supplemental/Tu_SpeeDe3DGS_Speedy_Deformable_CVPR_2026_supplemental.pdf",
    "https://github.com/tuallen/speede3dgs/tree/main"
  ],
  "raw_entry": "raw/ingest/2026-09-25-speede3dgs/",
  "topics": "[topics/3d-gaussian-splatting]",
  "entities": [],
  "claims": [],
  "questions": [],
  "tags": [
    "near-cvpr-2025",
    "3d-vision",
    "gaussian-splatting",
    "neural-rendering",
    "primary-source"
  ]
}