Yejun Zhang, Xinjue Wang, Zihan Wang, Esa Rahtu, Juho Kannala
GeoMix is a descriptor-free visual localization framework that significantly boosts accuracy by enhancing geometric discriminability at local, global, and training levels.
Descriptor-free visual localization avoids descriptor storage and privacy issues but suffers from low accuracy due to insufficient geometric discriminability. Existing methods underutilize local geometry cues, lack global context among keypoints, and overfit to a single keypoint detector.
GeoMix strengthens geometric discriminability in three ways: 1) Locally, it uses directional and distance-aware embeddings for richer neighborhood aggregation. 2) Globally, learnable context nodes aggregate scene-wide information via cross-attention to resolve ambiguities. 3) At the training level, Mix-Training learns representations across multiple keypoint detectors in a shared geometry-only space.
GeoMix sets a new state of the art among descriptor-free methods on MegaDepth, Cambridge Landmarks, 7Scenes, and Aachen Day-Night benchmarks. It reduces the 75th-percentile rotation error by 89% and translation error by up to 90% over the previous best, while generalizing zero-shot to unseen detectors and narrowing the gap to descriptor-based pipelines.