Based on this 360 video (credit to MTaka)
https://www.youtube.com/watch?v=Pvi5dHJobec
1. Download it from youtube using yt-dlp python package
2. Extracted images every 1 FPS totalling 567 of 8K equirectangular images (7680*3840) by using ffmpeg
3. Used modified colmap panorama_sfm.py. It takes 5 hours to process it with RTX3090 😂 . It worthed. All images is detected 🎉 . FYI, colmap panorama_sfm.py convert 1 equirectangular images into 12 pinhole images. Totaling 567*12=6804 images detected 👍🏼
4. Preprocessing (adjust contrast of the images) and create dense point cloud by Colmap
5. Trained with Brush commit 8ada1f7 (6 M splats, 100k iters, SH3)
7. Upload, rotate, and scale the splats, publish it with LOD in Supersplat.
8. Adds collision model in supersplat studio
A little bit quality degradation caused by compressed splat SOG. Alternative result from gsplat: https://superspl.at/scene/f4d32789