SfM-Free Support

[pfxuan]

Current 3DGS training pipeline is heavily relying on the SfM initialization, which introduces a significant overhead when scaling to large scenes. Technically, 3DGS should be capable of producing relevantly good results with a randomly or strategically initialized point cloud:

• https://oasisyang.github.io/colmap-free-3dgs/
• https://ku-cvlab.github.io/RAIN-GS/

While waiting for the release of COLMAP-Free 3DGS codebase, RAIN-GS appears to offer a similar or potentially simpler solution for decoupling the SfM step. This involves sparse-large-variance (SLV) initialization along with progressive gaussian low-pass filter control.

[pierotofy]

Interesting work for sure! The downside of colmap-free-2dgs is that the method relies on neural nets (DPT https://github.com/isl-org/DPT) for the monocular depth estimation step.