If the camera parameters are known, can we get dense point clouds from it?

[Vilour]

I have a COLMAP model with more than 100 images, can I use the COLMAP model as a fixed prior to get dense point cloud? I used to try with dust3r, but with too many images the performance degenerates, and there is out-of-memory problem. Can Spann3r preset pose and camera intrinsic like dust3r?

[lck666666]

same problem

[HengyiWang]

As far as I know, this preset pose and camera intrinsic happens during the optimization-based global alignment, where they freeze those preset params and optimize the rest of the camera params. In Spann3R, we do not have this alignment as the predictions are in a common coordinate system. Depending on your use case, you may want to do an alignment of the entire scene to your colmap model after obtaining dense pointmaps from Spann3R or directly obtain camera poses from predicted pointmaps using pnp-ransac as a coarse initialization for NVS.