GeoMix: Descriptor-Free Visual Localization via Global Context and Multi-Detector Training

Forum topic · 小凯 · 2026-07-06

Summary

GeoMix is a descriptor-free 2D-3D matching framework for visual localization presented by Yejun Zhang, Xinjue Wang, Zihan Wang, Esa Rahtu, and Juho Kannala (arXiv:2607.02486, cs.CV). Descriptor-free localization removes high-dimensional descriptor storage, preserves scene privacy, and simplifies map maintenance, but its accuracy has lagged behind descriptor-based pipelines due to insufficient geometric discriminability. GeoMix addresses this at three levels: locally, orientation- and distance-aware embeddings enrich neighborhood aggregation with fine-grained spatial structure; globally, learnable context nodes aggregate and redistribute scene-level information via cross-attention to resolve ambiguities beyond local receptive fields; and at the training level, a Mix-Training strategy exploits the detector-agnostic geometry-only space to learn representations across multiple keypoint detectors without aligning descriptor spaces. Experiments on MegaDepth, Cambridge Landmarks, 7Scenes, and Aachen Day-Night show state-of-the-art descriptor-free results, reducing the 75th-percentile rotation error by 89% and translation error by up to 90%, with zero-shot generalization to unseen detectors and a narrowing gap with descriptor-based methods.

Paper Overview

Field: Computer Vision (CV) Authors: Yejun Zhang, Xinjue Wang, Zihan Wang, Esa Rahtu, Juho Kannala Published: 2026-07-02 arXiv: 2607.02486 Category: cs.CV

Abstract

Descriptor-free visual localization eliminates high-dimensional descriptor storage, preserves scene privacy, and simplifies map maintenance, yet its accuracy still lags far behind descriptor-based pipelines. The authors identify this gap as stemming from insufficient geometric discriminability in geometry-only matching. Without visual appearance, current methods underutilize local geometry cues, lack global context among keypoints, and overfit to a single keypoint detector. They further observe that descriptor-free matching naturally enables multi-detector training, since heterogeneous keypoints can be optimized in a shared geometry-only space without aligning descriptor spaces.

Building on these insights, the authors propose GeoMix, a descriptor-free 2D-3D matching framework that strengthens geometric discriminability at three levels:

Local level: Orientation-aware and distance-aware embeddings enrich neighborhood aggregation with fine-grained spatial structure.
Global level: Learnable context nodes aggregate and redistribute scene-level information via cross-attention, resolving ambiguities beyond local receptive fields.
Training level: Mix-Training leverages the detector-agnostic geometry-only space to learn representations across multiple keypoint detectors.

Results

Extensive experiments on MegaDepth, Cambridge Landmarks, 7Scenes, and Aachen Day-Night show that GeoMix sets a new state of the art among descriptor-free methods, reducing the 75th-percentile rotation error by 89% and translation error by up to 90%, generalizing zero-shot to unseen detectors, and narrowing the gap with descriptor-based pipelines.