360° Gaussian Splatting Walkthroughs

Indoor scenes reconstructed from single 360° walking videos (Insta360 → equirectangular 3840×1920). Pipeline: frame extraction → spherical SfM (OpenSfM) → MVS dense initialization → 3DGS training.
The cafe scene is the density-optimized run (3M init points), reaching held-out PSNR 20.95 dB — first break past the 20 dB ceiling.

🌀 Interactive splat viewers (desktop with a decent GPU recommended)

☕ Cafe — density-optimizedNEW

2,269,510 gaussians · PSNR 20.95 dB · 73MB · drag to orbit, WASD to walk

🏢 Lobby + Corridor

1,154,312 gaussians · cleaned (floaters/streaks removed) · 37MB · drag to orbit, WASD to walk

📷 SfM reconstructions (point cloud + camera path, self-contained three.js)

☕ Cafe — SfMFULL 3M

full 3,000,000-point dense cloud + 230 camera frusta · 60MB, give it a few seconds

🏢 Lobby + Corridor — SfM

dense cloud + 149 camera frusta · self-contained three.js viewer

📈 Density scaling study (cafe scene)

How many initial points does 3DGS actually need? Same scene, only the MVS dense-init density changed.

400k → 1.5M gives +1.32 dB; 1.5M → 3M only +0.08 dB while doubling size and training time. densify/prune absorbs the extra points — the initial cloud just needs to be sufficient, not maximal. (The cafe scene above serves the 3M run.)