[{"data":1,"prerenderedAt":254},["ShallowReactive",2],{"work-gaussian-training":3,"works-order":198},{"id":4,"title":5,"attribution":6,"body":7,"caseOrder":6,"cover":148,"demo":6,"description":53,"evidenceStatus":152,"extension":153,"family":154,"featured":155,"flagship":155,"highlights":156,"links":160,"media":6,"meta":166,"metrics":167,"navigation":177,"order":139,"path":178,"period":179,"projectType":180,"projects":181,"role":183,"roleScope":184,"seo":185,"stack":186,"status":193,"stem":194,"summary":195,"topic":196,"track":180,"__hash__":197},"works\u002Fworks\u002Fgaussian-training.md","高斯泼溅重建与训练管线",null,{"type":8,"value":9,"toc":138},"minimark",[10,14,18,27,30,43,46,57,60,99,102,110,113,124,127],[11,12,13],"h2",{"id":13},"背景问题",[15,16,17],"p",{},"单相机视频与多相机阵列采集的数据，需要经过抽帧、去畸变、SfM、训练与导出等多步骤。手工脚本存在三个问题：结果不可复现、SfM 失败后仍会空耗 GPU 训练时间、中途崩溃无法续跑。",[15,19,20,21,26],{},"下游创作、编译与查看发布见",[22,23,25],"a",{"href":24},"\u002Fworks\u002Fscene-toolchain","Scene 离线三维创作与发布工具链","，本篇只覆盖训练侧。",[11,28,29],{"id":29},"约束",[31,32,33,37,40],"ul",{},[34,35,36],"li",{},"编排层尽量只依赖 Python 标准库，降低部署环境的依赖面。",[34,38,39],{},"Windows 与 Linux\u002FDocker 两套环境并存，配置需要保持键集合一致。",[34,41,42],{},"外部工具（COLMAP、FFmpeg、训练后端）版本要可锁定、可校验。",[11,44,45],{"id":45},"架构",[47,48,54],"pre",{"className":49,"code":51,"language":52,"meta":53},[50],"language-text","视频 \u002F 多路鱼眼 Rig\n  └─ FFmpeg 抽帧 + FishPoly 去畸变 + 位姿先验\n      └─ COLMAP SfM（rig 约束）──► sfm_validation.json（门禁）\n          └─ Brush 0.3.0（Windows）或 Nerfstudio splatfacto（Docker\u002FCUDA）\n              └─ 导出 .ply ──► splat-transform 生成 .compressed.ply \u002F .sog \u002F 独立查看器\n","text","",[55,56,51],"code",{"__ignoreMap":53},[11,58,59],{"id":59},"关键决策与取舍",[31,61,62,73,83,93],{},[34,63,64,68,69,72],{},[65,66,67],"strong",{},"状态机而非脚本","：prepare、sfm、convert、train、export、supersplat 六个阶段解耦。重跑某阶段时自动清理本阶段产物并使下游失效，",[55,70,71],{},"--from-stage"," 支持断点续跑。",[34,74,75,78,79,82],{},[65,76,77],{},"门禁前置","：训练前读取 ",[55,80,81],{},"sparse\u002F0","，检查有效图像数与注册率，不达标直接中止，避免算力空耗。",[34,84,85,88,89,92],{},[65,86,87],{},"工具版本锁定","：",[55,90,91],{},"tools\u002Fmanifest.toml"," 记录工具来源、版本与 SHA-256，bootstrap 时校验，不静默升级。",[34,94,95,98],{},[65,96,97],{},"训练后端二选一","：Windows 用 Brush（WebGPU\u002FVulkan），生产 Linux 用 splatfacto，两者共用上游数据契约。",[11,100,101],{"id":101},"结果",[31,103,104,107],{},[34,105,106],{},"同一份输入在两个后端之间的前后处理保持一致，失败点可以定位到具体阶段。",[34,108,109],{},"测试全部使用标准库 unittest，并校验两套配置的键集合一致。",[11,111,112],{"id":112},"局限与后续优化",[31,114,115,118,121],{},[34,116,117],{},"SfM 门禁为静态阈值（有效图像数 \u002F 注册率），对弱纹理与重复纹理场景仍依赖人工复核输入质量。",[34,119,120],{},"Brush 与 splatfacto 共用上游数据契约，但训练参数未完全对齐，跨后端结果存在不可比差异。",[34,122,123],{},"当前面向本机与单机 GPU 编排，未纳入分布式或队列优先级调度。",[11,125,126],{"id":126},"我的职责",[31,128,129,132,135],{},[34,130,131],{},"编排内核与状态机设计，阶段产物与级联失效规则。",[34,133,134],{},"SfM 门禁、rig 配置生成与鱼眼去畸变链路。",[34,136,137],{},"目录队列 Worker、事件流与 Docker 双 target 构建。",{"title":53,"searchDepth":139,"depth":139,"links":140},2,[141,142,143,144,145,146,147],{"id":13,"depth":139,"text":13},{"id":29,"depth":139,"text":29},{"id":45,"depth":139,"text":45},{"id":59,"depth":139,"text":59},{"id":101,"depth":139,"text":101},{"id":112,"depth":139,"text":112},{"id":126,"depth":139,"text":126},{"src":149,"alt":150,"caption":151},"\u002Fimages\u002Fportfolio\u002Fgaussian-training-cover.svg","3DGS 训练管线从输入、SfM 门禁到训练与导出的流程示意","训练管线结构示意（非运行截图）","项目资料与演示","md","Shanks Engine 重建训练管线",false,[157,158,159],"每次运行独立工作区，输入 SHA-256 清单与配置快照可追溯","多相机刚性约束与鱼眼去畸变，保持 rig 外参一致","目录队列 Worker 与 JSONL 事件流，便于后端编排",[161],{"label":162,"url":163,"kind":164,"note":165},"实验场（渲染产物）","\u002Flab","lab","管线产物的 3DGS 渲染演示",{},[168,171,174],{"value":169,"label":170},"6","幂等阶段（prepare → supersplat）",{"value":172,"label":173},"≥ 20 \u002F ≥ 0.85","SfM 门禁（有效图像数 \u002F 注册率）",{"value":175,"label":176},"2","GPU 训练后端（Brush \u002F splatfacto）",true,"\u002Fworks\u002Fgaussian-training","项目阶段记录","personal",[182],"gaussian-splatting-pipeline","独立设计与开发（编排内核、SfM 质量门禁、训练后端接入）","编排内核、SfM 质量门禁、训练后端接入",{"title":5,"description":53},[187,188,189,190,191,192],"Python 3.11","COLMAP 4.2","FFmpeg","Brush","Nerfstudio splatfacto","Docker","demo","works\u002Fgaussian-training","面向单相机与多相机刚性阵列的 3DGS 端到端流水线，具备状态机、质量门禁与断点续跑，减少手工脚本的不可复现问题。","spatial","RczXAdd1zRJeXTzMV5qGZWjRYlr7kg2tKMUcsumzGFA",[199,201,202,206,210,214,218,222,226,230,234,238,242,246,250],{"path":24,"title":25,"order":200},1,{"path":178,"title":5,"order":139},{"path":203,"title":204,"order":205},"\u002Fworks\u002Fmindcloudx-gen1","MindCloudX AI 一代 3DGS 平台",3,{"path":207,"title":208,"order":209},"\u002Fworks\u002Fexplorer-sdk","Explorer 多引擎渲染 SDK 与沉浸式浏览",4,{"path":211,"title":212,"order":213},"\u002Fworks\u002Ftorchcloud-gen2","TorchCloud 二代平台与混合场景编辑器",5,{"path":215,"title":216,"order":217},"\u002Fworks\u002Fdata-pipeline","3DGS 与 Mesh 数据转换管线",6,{"path":219,"title":220,"order":221},"\u002Fworks\u002Fmesh-pointcloud","大规模 Mesh 与点云可视化",7,{"path":223,"title":224,"order":225},"\u002Fworks\u002Fhunter-game","Hunter World：3DGS 真实场景多人寻宝竞速",8,{"path":227,"title":228,"order":229},"\u002Fworks\u002Fsitepulse-agent","智慧工地：空间可视化与 Agent 业务集成",9,{"path":231,"title":232,"order":233},"\u002Fworks\u002Fdrone-platform","无人机 GCS 与数据大屏交付",10,{"path":235,"title":236,"order":237},"\u002Fworks\u002Fcar-configurator","汽车 3D 配置器",11,{"path":239,"title":240,"order":241},"\u002Fworks\u002Fphotography-interaction","摄影器材 3D 交互展示页",12,{"path":243,"title":244,"order":245},"\u002Fworks\u002Fcloud-lims","Cloud LIMS 实验信息管理（Demo）",13,{"path":247,"title":248,"order":249},"\u002Fworks\u002Fcampus-platform","校园业务小程序与管理后台",14,{"path":251,"title":252,"order":253},"\u002Fworks\u002Fvisual-design","商业三维渲染与平面视觉设计",15,1792474478000]