DECO: Depth-Guided Co-Visibility Reasoning for Low-Altitude UAV Visual Localization

Yibin Ye; Xichao Teng; Shuo Chen; Xiaokai Song; Dongdong Guan; Qifeng Yu; Zhang Li

Source-linked AI summary

DECO: Depth-Guided Co-Visibility Reasoning for Low-Altitude UAV Visual Localization

Yibin Ye, Xichao Teng, Shuo Chen, Xiaokai Song, Dongdong Guan, Qifeng Yu, Zhang Li

arXiv:2608.22289v1Aug 23, 2026cs.CV

TL;DR

DECO addresses limitations in low-altitude UAV visual localization by using depth-guided co-visibility reasoning. Experiments report improved localization accuracy across different MDEs, while attitude-prior variants remain stronger overall.

Problem
Low-altitude UAV localization is limited by reference maps that omit or compress vertical structures and by methods focused mainly on local feature detection and description.
Method
DECO is a depth-guided co-visibility reasoning framework that uses local surface geometry to support UAV visual localization.
Results
DECO consistently improves localization accuracy across different MDEs, and monocular-depth geometry improves nearly all evaluated metrics when attitude priors are unavailable.
Takeaways & Limitations
Monocular-depth geometry can provide an effective substitute when attitude priors are unavailable, although the attitude-prior variant remains stronger overall.
Takeaways & Limitations
DECO still has several limitations, and its performance remains affected by factors identified in the paper’s discussion.

Abstract

from arXiv · show

Unmanned aerial vehicles (UAVs) increasingly require robust visual localization in GNSS-denied environments. A common solution estimates UAV poses by matching keypoints between UAV images and geo-tagged orthographic reference maps derived from satellite or aerial imagery, followed by Perspective-\(n\)-Point (PnP) pose solving. However, such reference maps mainly record top-down surfaces such as roofs and ground planes, while vertical structures such as facades and walls are often compressed or missing. Consequently, many visually distinctive keypoints in low-altitude UAV images have no valid counterparts in the reference map, leading to redundant matches and inaccurate pose estimation. To address this issue, we propose DECO, a DEpth-guided CO-visibility reasoning framework for low-altitude UAV visual localization. DECO uses monocular depth priors to infer local surface geometry and estimate co-visible regions between UAV images and the reference map. Based on this prior, a Geometry-Saliency Coupled Co-visibility Score is introduced to jointly consider geometric co-visibility and detector saliency for keypoint ranking. In this way, DECO retains keypoints that are both visually distinctive and geometrically co-visible, improving feature matching and PnP-based pose estimation. Extensive experiments demonstrate that DECO achieves superior localization performance and can be integrated with different depth models, feature detectors, and matchers. The source code will be available at https://github.com/UAV-AVL/DECO.

1 Introduction

Low-altitude oblique UAV images contain salient vertical structures that top-down orthographic maps represent poorly, creating a co-visibility gap and unreliable matches. DECO addresses this gap by using depth-guided co-visibility reasoning to rank keypoints before matching.

Motivation: Low-altitude oblique UAV images contain facades, walls, and object sides that top-down orthographic reference maps retain only weakly.This creates a co-visibility gap between the UAV image and reference map.
Motivation: Visually salient but non-co-visible regions can expand the search space, introduce unreliable correspondences, and weaken PnP pose estimation.
Research gap: Existing feature methods improve detection, description, or matching but generally do not explicitly reason about region-level co-visibility.
Research gap: Large viewpoint differences and complex remote-sensing scenes make overlap or semantic-consistency estimation difficult for low-altitude UAV localization.
DECO framework: DECO uses monocular depth to infer local surface geometry and estimate geometric co-visibility between UAV images and orthographic reference maps.
DECO framework: The GS-CoVis score combines geometric co-visibility and detector saliency to retain keypoints that are distinctive and geometrically consistent.
Reported outcome: Co-visibility reasoning before matching reduces redundant correspondences and improves the reliability of subsequent PnP-based pose estimation.
Reported outcome: DECO achieves superior localization performance and operates as a plug-and-play module across different depth models, detectors, and matchers.

2 Related Work

Cross-view localization methods span detector-based, detector-free, overlap-aware, semantic, and depth-guided approaches. DECO differs from prior depth-guided work by using depth-derived geometry for explicit co-visibility reasoning and keypoint ranking in UAV-to-map localization.

Feature matching paradigms: Detector-based pipelines detect and describe sparse keypoints independently, whereas detector-free methods directly predict dense or semi-dense correspondences.
Feature matching paradigms: Detector-free methods can be more robust under viewpoint changes but generally require higher computational and memory costs.
Feature matching paradigms: Detector-based methods remain attractive for onboard UAV localization because they are generally more efficient and easier to deploy.
UAV-specific challenges: In low-altitude UAV localization, independently detected keypoints may lie on vertical structures without valid orthographic-map counterparts, degrading pose estimation.
Prior-guided matching: Overlap, relative-scale, co-visibility, and semantic priors constrain matching toward regions likely to correspond across views.
Prior-guided matching: These prior-guided strategies face greater difficulty in low-altitude UAV scenes because of viewpoint differences, perspective distortion, and complex content.
Depth-guided matching: Earlier depth-guided methods use monocular depth for planar rectification, homography estimation, feature matching, or correspondence filtering.
Depth-guided matching: DECO instead uses depth-derived surface geometry for explicit cross-view co-visibility reasoning and co-visibility-aware keypoint ranking.

3 Methodology

DECO estimates co-visibility from monocular depth and uses it with detector saliency to rank UAV keypoints before matching and PnP pose estimation. The pipeline converts selected reference-map matches into 3D correspondences using a DSM and recovers the UAV location from the estimated camera pose.

3.2 Monocular-Depth-Based Geometric Co-visibility Prior: DECO predicts monocular depth and compares local surface normals with gravity to form a dense geometric co-visibility prior.The default implementation uses Depth Anything 3, while other monocular-depth estimators are evaluated separately.
3.3 Geometry-Saliency Coupled Co-visibility Score: The prior is coupled with detector saliency through the GS-CoVis score, ranking keypoints that are both geometrically co-visible and visually distinctive.Rank normalization reduces dependence on detector-specific confidence-score scales, while λ controls the balance between geometric co-visibility and saliency.
3.4 Co-Visibility-Aware Feature Matching and Pose Estimation: Top-ranked UAV keypoints are matched to reference-map keypoints, whose locations are lifted to 3D with the co-registered DSM for PnP pose estimation.RANSAC-based PnP estimates rotation and translation by minimizing reprojection error, after which the camera center is converted to geographic coordinates.
3.2 Monocular-Depth-Based Geometric Co-visibility Prior: The geometric score uses absolute cosine similarity between each local surface normal and camera-frame gravity; larger values indicate greater likely co-visibility with the orthographic map.The camera-frame gravity direction is obtained from the coarse roll and pitch attitude prior rather than directly constraining the final pose.

4 UAV Visual Localization Experiments

DECO is evaluated on AnyVisLoc and OrthoLoC under varied low-altitude imaging, domain, and matching conditions. Results show that co-visibility-aware keypoint selection improves geometric consistency and localization across pipelines, depth estimators, and benchmarks.

Datasets: AnyVisLoc emphasizes oblique views with vertical structures that are weakly represented in top-down reference maps, creating a co-visibility gap for evaluation.Most UAV images have pitch angles between 0° and 60°, and the test subset contains 2,492 image pairs from four urban scenes.
Generality analysis: DECO improves almost all detector–matcher pipelines on both benchmarks, with SP+LG T@3 gains of 4.4%, 10.2%, 2.8%, and 1.8% across AnyVisLoc scenes and 0.6% and 0.4% on OrthoLoC.The inlier ratio also improves in most cases, indicating more geometrically consistent matched 2D–3D correspondences.
Benchmark results: DECO’s improvements are more pronounced on AnyVisLoc and remain stable on OrthoLoC despite modality differences and domain shifts.The stronger AnyVisLoc gains are attributed to its more oblique imagery and richer vertical structures, while OrthoLoC results indicate robustness across viewing conditions and data sources.
Comparisons and analysis: DECO consistently improves the SP+LG baseline across most metrics and depth estimators, while outperforming semantic-prior and several geometry-aware alternatives in comparisons.Qualitative results associate DECO’s filtering of salient but non-co-visible structures with cleaner correspondences, more PnP inliers, and more accurate localization; combining DECO with LiftFeat also consistently improves localization.

5 Discussion

DECO’s sensitivity and robustness analyses show that balanced keypoint selection improves localization, while relaxed curation and uncertain orientation expose practical performance boundaries. A depth-derived dominant-plane strategy provides a fallback gravity estimate when attitude priors are unavailable, but depends on scene geometry and depth quality.

Sensitivity Analysis: Increasing K initially improves localization, then saturates as additional retained keypoints contribute less co-visible evidence and increase matching and verification cost.K = 1024 provides the reported trade-off between localization accuracy and computational efficiency.
Sensitivity Analysis: When K = 256, the retained set is too sparse to provide sufficient effective 2D–3D correspondences for robust matching and PnP-based pose estimation.
Sensitivity Analysis: Coupling detector saliency with geometric co-visibility enables DECO to retain keypoints that are both visually distinctive and geometrically consistent.The default λ = 1 lies in the stable and effective range reported in the analysis.
Reference-Map Curation and Localization Uncertainty: DECO generally retains an advantage over its baseline under relaxed spatial and orientation settings, although gains shrink when orientation uncertainty becomes large.Removing yaw constraints or enlarging the search region decreases localization accuracy and introduces challenges beyond co-visibility.
Gravity-Direction Prior: Plane-normal gravity estimation improves nearly all metrics over the SP+LG baseline but remains weaker overall than the attitude-prior variant.It provides a practical fallback when reliable attitude priors are unavailable, while its quality depends on monocular depth and scene layout.

6 Conclusion

DECO frames reliable UAV-to-map matching as a joint problem of detector saliency and geometric co-visibility. It shifts keypoint selection toward geometry-aware correspondence construction and improves localization across varied models while retaining plug-and-play integration.

DECO jointly considers detector saliency and geometric co-visibility to construct more reliable UAV-to-map correspondences.This replaces appearance-driven selection with geometry-aware correspondence construction.
DECO suppresses visually salient but geometrically non-co-visible structures before matching.
DECO consistently improves localization accuracy across different monocular depth estimation models, feature detectors, and feature matchers.
Compared with recent prior-guided matching methods, DECO more directly addresses the co-visibility gap in low-altitude UAV localization.
DECO remains affected by the accuracy and generalizability of its adopted MDE model, while adding MDE increases computational cost.Future work considers learning detectors under co-visible region supervision and tighter depth-semantic coupling.