DSM-Free UAV Visual Localization via Latent-Depth Variable Projection and Structure-Adaptive Coupling

Yibin Ye; Xichao Teng; Shuo Chen; Xiaokai Song; Yuli Sun; Qifeng Yu; Zhang Li

cs.CVOct 10, 2026

DSM-Free UAV Visual Localization via Latent-Depth Variable Projection and Structure-Adaptive Coupling

Authors: Yibin Ye, Xichao Teng, Shuo Chen, Xiaokai Song, Yuli Sun, Qifeng Yu, Zhang Li

Organizations: College of Aerospace Science and Engineering, National University of Defense Technology, Changsha 410073, China, and Hunan Provincial Key Laboratory of Image Measurement and Vision Navigation, Changsha 410073, China. · College of Electronic Science and Technology, National University of Defense Technology, Changsha 410073, China.

Abstract

Visual localization enables unmanned aerial vehicles (UAVs) to maintain absolute positioning in GNSS-denied environments. A common approach matches UAV images with geo-tagged orthographic maps and estimates camera pose using Perspective-n-Point (PnP) with elevations from Digital Surface Models (DSMs). However, accurate and up-to-date DSMs are often unavailable in practice. Orthographic-perspective geometry (OPG) enables DSM-free localization in scenes with varying elevations, yet existing methods typically estimate pose indirectly through intermediate matrix estimation and decomposition. To address this limitation, we propose Latent-Depth Variable Projection (LDVP), which treats the unknown scales along perspective viewing rays as latent depths. By analytically eliminating these depths and solving for the horizontal camera center in closed form given the rotation, LDVP reduces nonlinear estimation to a rotation-only problem, enabling direct recovery of metric horizontal position without assuming a planar scene. To improve localization robustness across different scene structures, we further introduce Structure-Adaptive Coupling (SAC), which uses relative scene structure inferred from monocular depth estimation to select LDVP when the matched points are not dominated by a single plane, and Planar-PnP otherwise. Our local observability analysis reveals how scene elevation variation affects the geometric constraints of OPG, providing a basis for this adaptive strategy. Extensive experiments show that LDVP consistently outperforms existing OPG-based DSM-free methods, while SAC further improves robustness across diverse scenes and substantially narrows the gap to DSM-based localization. Code will be released at https://github.com/UAV-AVL/LDVP-SAC.

Figures & tables

Fig. 1 : Overview of the proposed DSM-free UAV localization framework. LDVP directly estimates the metric horizontal camera position by analytically eliminating the latent depths. SAC uses relative scene structure inferred from monocular depth to adaptively select between LDVP and Planar-PnP, exploiting their complementary strengths across different scene structures.

Panel of Figure 2 — Fig. 2 : Synthetic evaluation under the smooth sinusoidal terrain model. The x-axis in (b) and (c) denotes matched-point relief ( $P_{90}-P_{10}$ , in meters), where $P_{90}$ and $P_{10}$ are the 90th and 10th percentiles of matched-point elevations.

Method	AnyVisLoc				OrthoLoC
Method	$A@1$ (%)	$A@3$ (%)	$A@5$ (%)	Med. (m)	$A@1$ (%)	$A@3$ (%)	$A@5$ (%)	Med. (m)
DSM-PnP	82.22	96.74	97.20	0.46	72.48	84.09	86.55	0.46
Planar-PnP	16.50	50.72	66.17	2.93	17.49	53.66	69.64	2.72
AH	28.01	54.87	64.99	2.44	19.16	35.37	41.68	11.36
OPE	41.37	70.76	80.91	1.37	36.73	59.66	69.28	1.83
LDVP (ours)	48.59	75.41	82.87	1.05	39.89	59.28	66.23	1.69

TABLE I : Localization results on AnyVisLoc and OrthoLoC using RoMa correspondences. Best and second-best DSM-free results are shown in bold and underline , respectively. DSM-PnP is excluded from the ranking.

Fig. 3 : Structural analysis of Planar-PnP, LDVP, and SAC on AnyVisLoc and OrthoLoC using RoMa correspondences. (a) Cumulative distribution of the true matched-point DSM relief, measured by

P_{90}-P_{10}

. (b) Fraction of queries for which Planar-PnP or LDVP yields the lower localization error in each relief interval. (c) Fraction of queries for which SAC selects Planar-PnP. (d) Agreement between the SAC-selected branch and the lower-error solver.

Fig. 4 : Scene-level localization results for Case 1 (AnyVisLoc Scene 07), a mixed urban–farmland environment. The upper region of the reference map (white polygon) contains pronounced urban height variations, whereas most of the remaining area is relatively flat farmland. The panels show the reference map, DSM, and localization results of DSM-PnP, Planar-PnP, OPE, AH, LDVP, and LDVP+SAC. Localization errors are color-coded using the $1$ m and $5$ m thresholds.

Panel of Figure 6 — Fig. 5 : Representative SAC decisions in Case 1 (AnyVisLoc Scene 07). Each example visualizes the UAV–reference matched points, reference DSM, monocular depth, recovered 3D structure and dominant-plane support, and the resulting localization estimates. DSM elevations are shown only for retrospective analysis and are not used by SAC.

Fig. 6 : Scene-level localization results for Case 2 (AnyVisLoc Scene 03), a dense urban environment with strong multi-level elevation structure. Planar-PnP degrades substantially in this regime, whereas LDVP and LDVP+SAC remain more stable.

Fig. 7 : Representative query from Case 2 (AnyVisLoc Scene 03). The matched points span strongly varying elevations, with a dominant-plane support ratio of $S_{1}=0.156$ ; SAC therefore selects LDVP.

Panel of Figure 9 — Fig. 8 : Representative SAC decisions on the OrthoLoC dataset. The top row shows two queries with weak dominant-plane support, for which SAC selects LDVP, while the bottom row shows two queries with strong dominant-plane support, for which SAC switches to Planar-PnP. Each example visualizes the UAV–reference matched points, reference DSM, monocular depth, recovered 3D structure and dominant-plane support, and the resulting localization estimates.

Dataset / $\tau_{S}$	SelP (%)	$\Delta A@1$	$\Delta A@3$	$\Delta A@5$
AnyVisLoc / 0.30	33.2	+0.74	+6.72	+6.99
AnyVisLoc / 0.40	23.3	+2.61	+7.21	+6.85
AnyVisLoc / 0.50	14.9	+2.76	+6.10	+5.49
AnyVisLoc / 0.60	8.3	+1.84	+3.84	+3.50
OrthoLoC / 0.30	57.02	-4.87	+4.72	+8.59
OrthoLoC / 0.40	33.47	-1.91	+4.52	+6.71

TABLE II : Sensitivity to the SAC consensus threshold using RoMa.

\Delta A@\tau

denotes the percentage-point change relative to LDVP; SelP denotes the percentage of queries assigned to Planar-PnP.

Matching method	Localization method	AnyVisLoc				OrthoLoC
Matching method	Localization method	$A@1$ (%)	$A@3$ (%)	$A@5$ (%)	Med. (m)	$A@1$ (%)	$A@3$ (%)	$A@5$ (%)	Med. (m)
SuperPoint+LightGlue	DSM-PnP	41.23	84.55	91.92	1.20	49.45	61.25	63.16	0.83
	Planar-PnP	7.98	33.20	49.99	4.93	9.21	34.92	47.82	4.52
	AH	8.07	23.87	33.28	11.37	10.77	23.50	29.70	23.50
	OPE	6.22	21.83	32.18	9.91	13.45	29.93	39.19	8.02
	LDVP (ours)	17.77	38.65	50.73	4.77	23.25	40.54	47.95	4.76

TABLE III : Localization results with different matching methods. Best and second-best DSM-free results are shown in bold and underline , respectively. DSM-PnP is excluded from the ranking.

Method	AnyVisLoc				OrthoLoC
Method	$A@1$	$A@3$	$A@5$	Med.	$A@1$	$A@3$	$A@5$	Med.
LDVP	48.59	75.41	82.87	1.05	39.89	59.28	66.23	1.69
LDVP+SAC (MoGe2-Aerial)	51.35	81.51	88.36	0.97	39.34	62.05	70.33	1.66
LDVP+SAC (MoGe2)	51.19	80.92	87.87	0.97	39.72	61.62	69.57	1.64
LDVP+SAC (Depth Anything 3)	51.25	81.64	88.50	0.97	38.53	62.49	71.34	1.70

TABLE IV : Ablation of the monocular depth model used by SAC with RoMa correspondences and

\tau_{S}=0.5

A@1

A@3

, and

A@5

are reported in percentages (%), and Med. in meters (m).

Robot Localization

Robot localization estimates a robot’s position and orientation from onboard sensors, maps, and observations of its surroundings.

See all 173 papers

DECO: Depth-Guided Co-Visibility Reasoning for Low-Altitude UAV Visual Localization
Aug 23, 2026Yibin Ye, Xichao Teng, Shuo Chen +4GNSS-Denied Localization Map-Based Localization
College of Aerospace Science and Engineering, National University of Defense Technology, Changsha 410073, China
RIM: A Retrieval-In-Matching Framework for Cross-Domain Global Visual Localization of UAVs
Jul 22, 2026Xin Li, Siyuan Duan, Shang Wang +3Cross-View Geo-Localization Visual Place Recognition
Key Laboratory of Spectral Imaging Technology CAS, Xi’an Institute of Optics and Precision Mechanics, Chinese Academy of Sciences, Xi’an, 710119, China · University of Chinese Academy of Sciences, Beijing, 100049, China
Unifying UAV Cross-View Geo-Localization via 3D Geometric Perception
Apr 2, 2026Haoyuan Li, Wen Yang, Fang Xu +4Cross-View Geo-Localization GNSS-Denied Localization
School of Electronic Information, Wuhan University, Wuhan 430072, China · School of Artificial Intelligence, Wuhan University, Wuhan 430072, China · Technology and Engineering Center for Space Utilization and the Key Laboratory of Space Utilization, Chinese Academy of Sciences, Beijing 100094, China +1

Aug 23, 2026J/K move · Enter open · S save

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

Yibin Ye, Xichao Teng, Shuo Chen, Xiaokai Song+3

College of Aerospace Science and Engineering, National University of Defense Technology, Changsha 410073, China

GNSS-Denied Localization Map-Based Localization Cross-View Geo-Localization

Open paper AskPDF arXiv

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.

Abstract

Robot Localization

Explore similar work

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

RIM: A Retrieval-In-Matching Framework for Cross-Domain Global Visual Localization of UAVs

Unifying UAV Cross-View Geo-Localization via 3D Geometric Perception