LLLMWIKI
ArticleResearch mapReading portalMetadata

Source note · Updated 2026-09-25

MoVieS有位姿单目视频的前馈动态高斯与运动建模

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的MoVieS从有相机内外参的单目动态视频前馈预测逐像素高斯和查询时间形变，联合NVS与点运动训练；0.93秒为已训练网络逐场推断，32张H20约五天的训练、相机估计与尺度校准另计（主文pp.297–300、补充p.1）。

问题、输入与机制

输入N帧RGB、内外参和时间，输出新位姿/新时间RGB、深度及三维跟踪；未给位姿的视频需外部估计。首帧坐标系逐像素建立高斯，VGGT预训练骨干加相机射线/token及时间token，深度/外观/运动DPT头分别回归，AdaLN查询时间给出位置与外观形变；有深度/点轨迹标注时联合深度、渲染与运动损失（§3 pp.297–298 Eqs.1–7）。RE10K训练尺度需外部Video Depth Anything、Depth Pro和MegaSaM重估（补充p.1）。

证据与边界

  • 主文p.300 Table2：DyCheck可见掩码mPSNR18.46/mLPIPS.3094对MoSca18.24/.3698，NVIDIA PSNR19.16/LPIPS.3152却低于MoSca21.45/.2653；后者45分钟逐场优化，输入动态分割mask设定亦影响比较。RE10K静态阶段PSNR27.60高于完整动态版26.98。
  • 主文p.300 Table3：Aria Digital Twin EPE3D .2153对SpatialTracker .5413；DriveTrack δ0.10 79.87%低于SpatialTracker80.67%。二维基线额外接Video Depth Anything，所有预测三维点中位范数归一，真实绝对尺度另验。
  • 主文pp.301–302 Tables4–6：无相机条件/仅射线/仅token/组合，静态PSNR分别25.56/25.81/26.81/27.60；去运动监督Aria EPE .7938对联合.2153。独立的运动分割、scene flow仅有Fig.5和补充Fig.S.3–S.4定性图，无量化GT支持。
  • 主文p.299、补充pp.1–2：八训练集含真实/合成及伪标注，32×H20约五天，动态阶段渐进5→13输入帧；峰显存和包含相机求解的总延迟未报告，仓库MIT但部分依赖/数据有非商业许可。无本机GPU复现。

相关页面

3D Gaussian Splatting: Representation, Optimization, Geometry, and Dynamic Scenes；SpeeDe3DGS：动态高斯时间剪枝与运动分组。本工作跨场预训练推断与SpeeDe3DGS逐场剪枝的训练/推断分母须分别记账。逐页矩阵见raw/ingest/2026-09-25-movies/analysis.md。

Metadata
{
  "id": "2026-09-25-movies",
  "type": "source",
  "title": "MoVieS：有位姿单目视频的前馈动态高斯与运动建模",
  "status": "reviewed",
  "note_status": "reviewed",
  "created": "2026-09-25",
  "updated": "2026-09-25",
  "source_type": "paper",
  "venue": "CVPR 2026",
  "authors": "[Chenguo Lin, Yuchen Lin, Panwang Pan, Yifan Yu, Tao Hu, Honglei Yan, Katerina Fragkiadaki, Yadong Mu]",
  "published_at": "2026-06-01",
  "ingested_at": "2026-09-25",
  "canonical_links": [
    "https://openaccess.thecvf.com/content/CVPR2026/html/Lin_MoVieS_Motion-Aware_4D_Dynamic_View_Synthesis_in_One_Second_CVPR_2026_paper.html",
    "https://openaccess.thecvf.com/content/CVPR2026/papers/Lin_MoVieS_Motion-Aware_4D_Dynamic_View_Synthesis_in_One_Second_CVPR_2026_paper.pdf",
    "https://openaccess.thecvf.com/content/CVPR2026/supplemental/Lin_MoVieS_Motion-Aware_4D_CVPR_2026_supplemental.pdf",
    "https://github.com/chenguolin/MoVieS"
  ],
  "raw_entry": "raw/ingest/2026-09-25-movies/",
  "topics": "[topics/3d-gaussian-splatting]",
  "entities": [],
  "claims": [],
  "questions": [],
  "tags": [
    "near-cvpr-2025",
    "3d-vision",
    "gaussian-splatting",
    "neural-rendering",
    "primary-source"
  ]
}