gaussian-splatting-cuda
gaussian-splatting-cuda copied to clipboard
Implement Visual SLAM
Moonshot -> The goal is to remove the Colmap dependency eventually. There is no plan yet how to go about it. However, at least we should get some pose priors to feed that into the global BA.
Other idea: https://x.com/hankyang94/status/1922736839410241882
Something like https://poiw.github.io/MODP/index.html or https://repo-sam.inria.fr/nerphys/on-the-fly-nvs/ to create initial poses, applying a round global BA and retraining might be also a quick way to outperform Colmap!
That would require to also port some mono-depth estimation model.