LLLMWIKI
ArticleResearch mapReading portalMetadata

Source note · Updated 2026-09-25

GaussianDWM驾驶语言高斯连接理解与生成

Read the paper

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

English translation pending — this page shows the Chinese original.

一句话结论

CVPR 2026以含语言特征的已构建驾驶场景高斯采样给Qwen3理解任务，再把语言知识与稀疏图像/深度条件用于生成；平均理解及近距离FID有改进，规划/描述及大平移视频指标存在反证（主文pp.10658–10663 Tables1–4，补充§B–E）。

输入、建场与两端机制

nuScenes六相机RGB、LiDAR累积点、相机/文字输入；每场先建3DGS并筛低质量场，SAM/CLIP生成512维语言特征，用逐场AE压到3维，再按LangSplat学语言高斯。位置和高斯属性投影到4096维token，场景QA用uniform+top-k选4096粒，定位QA增加query相似性筛选；Qwen3-8B先5k对齐、再30k LoRA。生成端以LLM高层文本和点云投影的低层条件，用冻结VAE、UNet流匹配产生RGB-D空间移位或未来视频；训练由224×400逐步升至424×800（主文§3–4、补充§B/C）。LiDAR、分割/CLIP、逐场训练与多个基础模型是进入推断前的重要投入。

主表与反证

NuInteract约150万多视QA标注；作者由nuScenes建800场高斯并过滤模糊、低重叠/PSNR场（补充§B p.1）。p.10661 Table1作者平均59.23对DriveMonkey52.12，2D定位mAP34.95对19.47；然而描述CIDEr78.72对79.79、规划准确80.95对82.64，3D VG Pr50.66对51.90。Table2同模型去高斯均值53.32，加随机高斯55.21，uniform+top-k58.93，另加相似性59.23；最后增量.30。主文Table1 CIDEr78.72与Table2完整78.82不一致，保留独立表值。nuScenes平移±1m作者FID8.36/FVD44.50优于DiST-S10.12/45.14；±4m作者18.81/116.40弱于DiST-S17.57/105.29（p.10663 Table3）。补充Table3在30场子集看世界知识+高斯±4m FVD150.05，比无两条件137.31更差，分母和主表不同。没有相机闭环安全/真实几何验证。

成本与证据边界

理解端16×A100，生成端多分辨率/多阶段及联合训练；从每场LiDAR建场、SAM/CLIP/AE、LLM、生成的总卡时、峰显存与端到端延迟未报告（主文§4.2、补充§B/C）。补充p.4 Table1采样token数量512方案平均58.26，高于普通4096方案56.92但交叉注意配置不同。公开仓库有代码目录，根目录无LICENSE，复用许可待核；无GPU复现。不同任务输入与评测指标应分别与LangSplat：高斯语言特征与三级语义查询、Diff4Splat:Repurposing Video Diffusion Models for Dynamic Scene Generation对照。详审raw/ingest/2026-09-25-gaussiandwm/analysis.md；3D Gaussian Splatting: Representation, Optimization, Geometry, and Dynamic Scenes。

Metadata
{
  "id": "2026-09-25-gaussiandwm",
  "type": "source",
  "title": "GaussianDWM：驾驶语言高斯连接理解与生成",
  "status": "reviewed",
  "note_status": "reviewed",
  "created": "2026-09-25",
  "updated": "2026-09-25",
  "source_type": "paper",
  "venue": "CVPR 2026",
  "authors": "[Tianchen Deng, Xuefeng Chen, Yi Chen, Qu Chen, Yuyao Xu, Lijin Yang, Le Xu, Yu Zhang, Bo Zhang, Wuxiong Huang, Hesheng Wang]",
  "published_at": "2026-06-01",
  "ingested_at": "2026-09-25",
  "canonical_links": [
    "https://openaccess.thecvf.com/content/CVPR2026/html/Deng_GaussianDWM_3D_Gaussian_Driving_World_Model_for_Unified_Scene_Understanding_CVPR_2026_paper.html",
    "https://openaccess.thecvf.com/content/CVPR2026/papers/Deng_GaussianDWM_3D_Gaussian_Driving_World_Model_for_Unified_Scene_Understanding_CVPR_2026_paper.pdf",
    "https://openaccess.thecvf.com/content/CVPR2026/supplemental/Deng_GaussianDWM_3D_Gaussian_CVPR_2026_supplemental.pdf",
    "https://github.com/dtc111111/GaussianDWM"
  ],
  "raw_entry": "raw/ingest/2026-09-25-gaussiandwm/",
  "topics": "[topics/3d-gaussian-splatting]",
  "entities": [],
  "claims": [],
  "questions": [],
  "tags": [
    "near-cvpr-2025",
    "3d-vision",
    "vision-language",
    "gaussian-splatting",
    "primary-source"
  ]
}