LLLMWIKI
正文研究地图阅读入口元数据

来源笔记 · 更新于 2026-09-25

Changes in Real Time在线多视场景变化检测与高斯更新

阅读原论文

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

一句话结论

离线建参考3DGS后，用PnP求新巡视帧位姿、融合像素/特征差异并持续更新变化高斯；作者在PASLCD在线mIoU .486/F1 .638、11.2FPS，巡视完成后选择性更新约42秒，这两个时间终点和参考建图成本需分开（主文pp.32251–32253 Tables1/4）。

输入、机制与基线

参考期SfM相机和Speedy-Splat建静态场；新巡视仅输入RGB、未知外参但需与参考视角和XFeat特征匹配，PnP/RANSAC+GPU miniBA估姿态。沿新姿态渲染参考图，计算L1+D-SSIM与SAM2-Tiny特征差，持久变化表示Rchange借自监督损失在已见多视角持续优化；巡视后仅对变化掩码重建图元并限制增密，再全局调色/修边（主文pp.32248–32250 Fig.3/Eqs.1–3）。假设参考图覆盖足够、巡视时场景近静止且相机可重新定位。与TT-Occ的语义体素输出不同，这里终点是相关变化区域二值掩码及后续可渲染的地图。

PASLCD 10场×两光照共20实例。主文p.32251 Table1在线ours mIoU .486/F1 .638/11.2FPS，同在线CS+CYWS2D .243/.360/8.2FPS；离线ours .552/.694/156秒、MV3DCD .478/.628/479秒。在线与离线可访问帧集合不同，离线静态pairwise基线有对齐便利，实验为作者设定内比较。p.32251 Table2 单帧89.60ms中多视融合58.17ms占64.9%；环境为RTX4090，1008×560，参考图构建没有纳入FPS。p.32253 Table4 PASLCD更新ours PSNR23.70/SSIM.7868/LPIPS.2491/42秒，重新训练3DGS 22.21/.7558/.2426/550秒；CL-Splats五小场ours 30.54/.9356/.1256/39秒 vs3DGS30.31/.9319/.1178/364秒。LPIPS两个集均稍差，保留画质反证。

消融、边界与复核

主文p.32252 Table3去L1 mIoU .320、完整 .486，改回阈值交集 .350；只留单一线索或去面积正则使优化坍塌。p.32253 Table5只做选择性重建28秒但PSNR19.89，合并全局步骤36秒/23.70，该消融与Table4 42秒按独立口径记录。补充p.1 Table6定位以COLMAP估计值作伪真值，约30ms/帧，未有测量GT；补充p.2 §7的XFeat强平面内旋转、低于VGA的输入密度和极大场景渲染是失败边界，地图无显式历史恢复。仓库GitHub API license=null、根目录无LICENSE，授权待核；论文未给参考建图端到端和峰显存。需按scene聚类验证F1、LPIPS、强光照/旋转鲁棒性及在线p95。未做GPU复现。审计见raw/ingest/2026-09-25-changes-in-real-time/analysis.md及三维高斯泼溅：表示、优化、几何与动态任务。

元数据
{
  "id": "2026-09-25-changes-in-real-time",
  "type": "source",
  "title": "Changes in Real Time：在线多视场景变化检测与高斯更新",
  "status": "reviewed",
  "note_status": "reviewed",
  "created": "2026-09-25",
  "updated": "2026-09-25",
  "source_type": "paper",
  "venue": "CVPR 2026",
  "authors": "[Chamuditha Jayanga Galappaththige, Jason Lai, Lloyd Windrim, Donald Dansereau, Niko Suenderhauf, Dimity Miller]",
  "published_at": "2026-06-01",
  "ingested_at": "2026-09-25",
  "canonical_links": [
    "https://openaccess.thecvf.com/content/CVPR2026/html/Galappaththige_Changes_in_Real_Time_Online_Scene_Change_Detection_with_Multi-View_CVPR_2026_paper.html",
    "https://openaccess.thecvf.com/content/CVPR2026/papers/Galappaththige_Changes_in_Real_Time_Online_Scene_Change_Detection_with_Multi-View_CVPR_2026_paper.pdf",
    "https://openaccess.thecvf.com/content/CVPR2026/supplemental/Galappaththige_Changes_in_Real_CVPR_2026_supplemental.pdf",
    "https://github.com/Chumsy0725/O-SCD"
  ],
  "raw_entry": "raw/ingest/2026-09-25-changes-in-real-time/",
  "topics": "[topics/3d-gaussian-splatting]",
  "entities": [],
  "claims": [],
  "questions": [],
  "tags": [
    "near-cvpr-2025",
    "3d-vision",
    "gaussian-splatting",
    "segmentation",
    "primary-source"
  ]
}