Test run of using 360 video to create a splat of a small weir.
Workflow:
1) Recorded with GoPro Max V2 @ 5.7K, timelapse video @ 2 fps with 1.2 m selfie stick.
2a) Stechdrive 1.16.0 for frame extraction (174 frames), “person” + sky + stitching masking (SAM 3), 5 face‑cube extraction (no top face) and COLMAP (exhaustive + incremental).
2b) Stechdrive again to generate mask without sky mask, for training only.
3) LightFeld Studio 0.5.2 nightly (fixes mask support) — MRNF — 130,000 iterations (less would've been fine) and PPISP @ 125,000 iters (cut from 5M to ~2.5M splats without noticeable visual degradation; too late to allow many new splats to be generated but nice filesize decrease). SH 3.
4) Supersplat for rotation + scaling (5 tiles ~1.5 m) + clipping + camera path. No cleanup.
Happy with:
- Very nice result for minimal recording effort (few‑minute walk).
- Love how the reflections in the water turned out.
- Plants are rendered with very acceptable quality; splatting is well suited to making it look OK when they are a bit fuzzy. Unfortunately man‑made objects don’t have such lenience.
To do:
- Record @ 8K (oops)
- Improve errors likely caused by stitching (the thin elements are wobbly). spheresfm? fisheye?
- Experiment with AprilTags for automatic scaling and easier COLMAP.
- Do something about reflection artifacts near shoreline (likely caused by holes in the splat).
- Pose optimisation?
- Figure out how to upright the sparse model in LightFeld Studio so training looks better (with the 3 point align, but it's very fiddly).