We built DRS-VPT to estimate a camera's 6-DoF pose in a 3D map from a single photo and a 3D point cloud, without an initial guess or known camera intrinsics.
The interesting part is generalisation: the same model works across driving LiDAR scans (KITTI, nuScenes), iPhone photos in indoor scans, and SFM maps.
We provide the code, model and weights, but you can also try it out at https://huggingface.co/spaces/fulkast/drs-vpt-demo
Happy to answer questions about the method or its limitations!