Recently I get stumbled by ChatGPT about best workflow for creating 3DGS from 360 images. He said the best workflow by using raw dual fisheye images instead of colmap `panorama_sfm` (converting equirectangular images into 12 faces or 6 cubical faces).
In order of experimentation, I will share some dataset, it originally came from "room3" dataset from *Full Circle* paper.
I also train it with 🖌️ Brush & Lichtfeld Studio for both (raw dual fisheye & colmap PanoSFM 12 faces). Feel free to train again with others software:
https://puprtes-my.sharepoint.com/:f:/g/personal/m_ichsan_pu_go_id/IgAIjWyGJdV9RJQR5AEZmdqgAf0PLccyIWZP5oCAIkhIGoc?e=uIOGUF