Copy & Paste the World with Gaussian Splatting
summary
Gaussian splatting converts phone, drone or 360 footage into explorable 3D scenes made of ellipsoid “splats,” often captured in 5–10 minutes and streamed or edited for VR and web. Production details: LiDAR pipelines produce about 10 GB per km and trains scan every 6–8 weeks; SNCF feeds LiDAR into Nvidia Omniverse + CloudXR to stream meshes for collision checks and vegetation reports, which works for planning. Formats and tools: Niantic Scaniverse, SBZ/PLY, SuperSplat editor, Marble and consumer apps, and the 3js main branch supports Gaussian splats as of last week.
AI-written summary · about
chapters (23)
- 0:00Introduction and speaker background
- 1:16Why copy and paste the world in 3D
- 2:00Traditional 3D capture methods
- 3:45Lidar scans and point clouds
- 4:56Displaying 3D spatial data
- 6:43Industries that benefit from 3D capture
- 7:46Railway infrastructure use case
- 9:30Digital twin platform for lidar data
- 13:39What is Gaussian Splatting
- 15:25Gaussian Splatting data structure
- 16:33Gaussian Splatting examples
- 18:22Capturing large areas with 360 cameras
- 19:10Drone-based Gaussian Splatting
- 20:13Real estate applications
- 21:46Editing and combining Gaussian Splats
- 22:50Tools for viewing and editing splats
- 23:54AI-generated Gaussian Splats
- 25:45Industries using Gaussian Splatting
- 27:53How to create Gaussian Splats yourself
- 29:57Public Gaussian Splats and sharing
- 31:10Enterprise Gaussian Splatting workflow
- 32:53Complex environment captures
- 33:55Q&A
