mickey

Why we included this project

MicKey is the reference implementation of a CVPR 2024 paper that estimates metric, scale-aware relative pose between two images without any prior 3D map. Rather than matching keypoints in pixel space and recovering scale separately, it regresses keypoint positions directly in camera space and learns metric correspondences end-to-end through differentiable pose optimization, needing only image pairs and their ground-truth poses for training. That makes it a natural fit for the Map-free benchmark, where a single reference photo stands in for a full 3D map, and for anyone working on augmented reality or visual relocalization who wants a reproducible baseline against classical RANSAC-based methods. The repo includes evaluation scripts and pretrained models, so you can run the method on your own image pairs. One caveat: the license is non-commercial, so treat it as research and teaching material rather than something to ship in a product.