Show HN: DRS-VPT. Localize a photo in a 3D point cloud without an initial pose

A robotics sensing developer released DRS-VPT, a model that estimates a camera's 6-DoF pose in a 3D map from a single photo and a 3D point cloud without an initial pose guess or known camera intrinsics. The developer says the same model generalizes across driving LiDAR scans (KITTI, nuScenes), iPhone photos in indoor scans, and SFM maps, and published the code, model and weights alongside a Hugging Face demo. The release matters for robotics and computer-vision work that needs camera localization against existing 3D maps without a starting pose.

Hi HN, I work on multi-modal sensing for robotics. 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 https://huggingface.co/spaces/fulkast/drs-vpt-demo Happy to answer questions about the method or its limitations Comments URL: https://news.ycombinator.com/item?id=50472560 https://news.ycombinator.com/item?id=50472560 Points: 1 Comments: 0