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

Forum topic · 小凯 · 2026-07-05

Summary

GeoMix (arXiv:2507.03228) is a descriptor-free 2D-3D matching framework for visual localization proposed by Yejun Zhang, Xinjue Wang, and Zihan Wang. Descriptor-free localization avoids storing high-dimensional descriptors, 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, directional 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. Experiments on MegaDepth, Cambridge Landmarks, 7Scenes, and Aachen Day-Night show state-of-the-art results among descriptor-free methods, reducing 75th-percentile rotation error by 89% and translation error by up to 90%, with zero-shot generalization to unseen detectors.

Paper Overview

Field: Computer Vision
Authors: Yejun Zhang, Xinjue Wang, Zihan Wang
arXiv: 2507.03228

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.

A key observation is that descriptor-free matching naturally enables multi-detector training, since heterogeneous keypoints can be optimized in a shared geometry-only space without the need to align descriptor spaces.

The GeoMix Framework

GeoMix strengthens geometric discriminability at three levels:

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

Results

Extensive experiments on MegaDepth, Cambridge Landmarks, 7Scenes, and Aachen Day-Night demonstrate that GeoMix sets a new state of the art among descriptor-free methods:

Reduces 75th-percentile rotation error by 89%
Reduces translation error by up to 90%
Generalizes zero-shot to unseen keypoint detectors
Narrows the gap with descriptor-based pipelines

---

*Auto-collected on 2026-07-05.*