LLLMWIKI
正文研究地图阅读入口元数据

来源笔记 · 更新于 2026-09-25

Off The Grid按图像局部结构预测亚像素高斯

阅读原论文

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

一句话结论

CVPR 2026把前馈3DGS中每个像素或每个体素固定输出图元的假设改成每patch软坐标检测、按熵分配16/32/64图元，改善作者同表新视角图像质量与图元密度；pose-free依赖预训练VGGT几何和评测期二次相机推断（主文§3–4 pp.11758–11762）。

输入与机制

2–12张静态RGB，无外部给定相机；VGGT-1B经DINOv2和跨视transformer估计内外参与深度，U-Net从patch特征生成P张heatmap，空间softmax坐标期望形成连续中心；深度/像素/特征双线性取样、MLP解码尺度/方向/opacity/confidence，多图变换到统一场景，测试删除αc<.1图元（pp.11758–11760 Fig.2–3/Eq.1）。颜色直接采输入像素，无球谐视角色。输入视图自渲染+跨视渲染、深度/法线一致和VGGT教师保持几何，未经GT 3D标注但需要冻结教师；训练场景的留出目标帧仅用于评测。

主表与反证

  • 六集合平均3/6/9/12视，作者PSNR21.21/SSIM.6470/LPIPS.3532，#G/输入像素.1431；AnySplat17.71/.5075/.3937与.8141，DA3-Giant18.83/.5428/.3834与1。作者所称“约七倍少”以像素密度为分母，模型/相机底座不相同（p.11761 Table1）。
  • GT深度与相机轴：Charge+SCRREAM平均AbsRel作者.1433 vs DA3-Giant.1339；相机AUC@30作者.9278 vs DA3-Giant.9339。作者FoV角误差.96最佳；NVS收益不表示所有GT几何轴领先（p.11763 Table2）。
  • 局部对照T&T+m360平均PSNR/LPIPS：full18.63/.3796；pixel-aligned18.33/.4341；pixel+offset18.33/.4744。DL3DV消融去confidence17.80/.4459、去adaptive18.72/.3898、全合19.09/.3379；对照场景和指标分母不同（p.11763 Tables3–4）。

局限、成本、追踪

补充p.1指出反光、天空/背景过近、天花板深度失真；主文§5 p.11763指出未见区域空洞、人像与光照变化退化。单GPU 140GB VRAM训练，训练总卡时及峰显存未报告；测试目标相机通过context及context+target两次前馈并求共同尺度，完整相机/渲染耗时需实测（p.11761–11762）。项目仓库MIT，代码/权重线索见raw/ingest/2026-09-25-off-the-grid/links.yaml；未做GPU复现。逐页版本、数据划分和原表见raw/ingest/2026-09-25-off-the-grid/analysis.md。三维高斯泼溅：表示、优化、几何与动态任务；Uni3R：无外参多视图的几何与语言语义联合高斯；EcoSplat：按目标图元数条件化前馈高斯。

元数据
{
  "id": "2026-09-25-off-the-grid",
  "type": "source",
  "title": "Off The Grid：按图像局部结构预测亚像素高斯",
  "status": "reviewed",
  "note_status": "reviewed",
  "created": "2026-09-25",
  "updated": "2026-09-25",
  "source_type": "paper",
  "venue": "CVPR 2026",
  "authors": "[Arthur Moreau, Richard Shaw, Michal Nazarczuk, Jisu Shin, Thomas Tanay, Zhensong Zhang, Songcen Xu, Eduardo Pérez-Pellitero]",
  "published_at": "2026-06-01",
  "ingested_at": "2026-09-25",
  "canonical_links": [
    "https://openaccess.thecvf.com/content/CVPR2026/html/Moreau_Off_The_Grid_Detection_of_Primitives_for_Feed-Forward_3D_Gaussian_CVPR_2026_paper.html",
    "https://openaccess.thecvf.com/content/CVPR2026/papers/Moreau_Off_The_Grid_Detection_of_Primitives_for_Feed-Forward_3D_Gaussian_CVPR_2026_paper.pdf",
    "https://openaccess.thecvf.com/content/CVPR2026/supplemental/Moreau_Off_The_Grid_CVPR_2026_supplemental.zip",
    "https://github.com/OffTheGridGS/OffTheGrid"
  ],
  "raw_entry": "raw/ingest/2026-09-25-off-the-grid/",
  "topics": "[topics/3d-gaussian-splatting]",
  "entities": [],
  "claims": [],
  "questions": [],
  "tags": [
    "near-cvpr-2025",
    "3d-vision",
    "gaussian-splatting",
    "neural-rendering",
    "primary-source"
  ]
}