Are you as sick and tired as I am of trying to extract camera poses for your NeRF's? Well if so I've got a treat for you: A fully deployed, GPU-accelerated Colmap endpoint that can turn your video into a fully-configured Colmap workspace!
Installs Docker and Cog, pulls image from Replicate, and runs on a sample video from my trip to the Louvre.
utils/Quickstart.sh
- Create Replicate endpoint
- Plug-and-play with nerfstudio
- Build tests
- Decrease run time (currently ~30 minutes). COLMAP options, downsample earlier?
- Optimize .zip file size (currently ~1GB). Remove extra images?
- Add Image and Insta360 input options
- COLMAP for their amazing toolset that is used in EVERY NeRF deployment.
- nerfstudio for their Python scripts.
- svox2 for their neural network-free radiance field utilities.