My first experimental splat, based on a modified COLMAP panorama_sfm.py script. Created from frames of a 2-minute 8K 24fps Dji Osmo 360 video, recorded in one loop around a church roughly 1 meter above my head.
Trained using Nerfstudio's splatfacto in sections—4 sides separately, with about 75 equirectangular frames per side. Each equirectangular frame was reframed into 12 pinhole views (2016x2016 px) for COLMAP SfM, and then trimmed to 4 views for splat training.
It required a lot of manual work to align the separate splats using the Snap plugin. Next time, I plan to align all valid cameras directly in COLMAP—which isn't fast or easy either—and train the 4 side subsets to avoid tedious manual alignment. I wouldn't recommend aligning more than two splats with this plugin to avoid issues.
Snap plugin: https://github.com/VFX-Soup/supersplat-snap
Google Maps: https://maps.app.goo.gl/uzn1BY8eydSoa8En7