SatFix: Absolute Visual Localization of UAVs in Satellite Maps from a Single Oblique Image

Jiarui Zeng; Kun Shi; Chiman Vong; Zhedong Zheng

cs.CVOct 8, 2026

SatFix: Absolute Visual Localization of UAVs in Satellite Maps from a Single Oblique Image

Authors: Jiarui Zeng, Kun Shi, Chiman Vong, Zhedong Zheng

Organizations: Department of Computer and Information Science, Faculty of Information Science and Computing, University of Macau, Macao SAR, China · Institute of Artificial Intelligence and Brain Sciences, University of Macau, Macao SAR, China

Abstract

We study absolute metric UAV localization within a provided geo-referenced satellite region, recovering continuous map position and viewing heading from a single oblique image or a short multi-view clip. Existing cross-view geo-localization methods retrieve the most similar satellite tile from a gallery and report Recall@K, but retrieval depends on gallery sampling, provides no heading estimate, and returns a tile index rather than a continuous coordinate. We propose SatFix, a feed-forward UAV--satellite localization framework built on VGGT- $Ω$ . Satellite-grid features act as queries that aggregate UAV visual evidence, and two lightweight heads regress a 3-DoF pose in the satellite-map frame: continuous 2D position and heading. SatFix requires no explicit 3D map, rendered bird's-eye image, auxiliary sensor, or test-time pose alignment. A single model supports both single- and multi-view inputs, with trajectory constraints used during multi-view training. For metric evaluation, we introduce University-Metric, where satellite imagery is re-collected over a region up to 10.7 $\times$ longer on a side (about 114 $\times$ the ground area) than the original University-1652 tiles, with continuous position and heading labels for the original UAV tours. With one UAV view, SatFix localizes 52.08% of test frames within 50 m and 17.34% within 10 m, with median position and heading errors of 45.66 m and $20.81^\circ$ , respectively. Inference takes under 0.1 s per single-view query on an NVIDIA RTX 4090. With nine UAV views, the median position error falls to 21.96 m and the median heading error to $8.73^\circ$ . Compared with a fine-tuned VGGT- $Ω$ baseline, SatFix reduces median position error by 34.0% and nine-view median heading error from $25.43^\circ$ to $8.73^\circ$ .

Figures & tables

Fig. 1: Task motivation. When GNSS is unavailable, e.g. , in urban canyons or under dense canopy, absolute metric UAV localization can use onboard vision. Given a single UAV view or a short clip of multiple views as input, the model outputs a position and a heading on a geo-referenced satellite map. Top: the UAV clip input and the satellite map. Bottom: ground-truth ( blue ), VGGT- $\Omega$ zero-shot ( red ), and SatFix ( cyan ) trajectories.

Benchmark	Input	# UAV Imgs.	Sat. Extent (m)	UAV Viewpoint	Sat. Ref.	Loc. Label	Task
University-1652 [ 1 ]	Single	89.2K	118	Oblique	Gallery	ID	Retrieval
SUES-200 [ 7 ]	Single	40.0K	N/A	Oblique	Gallery	ID	Retrieval
DenseUAV [ 2 ]	Single	9.1K	120	Nadir	Gallery	Coord.	Retrieval
UAV-VisLoc [ 4 ]	Single	6.7K	11000	Nadir	Regional map	Coord.+Heading	Localization
GeoVINS [ 5 ]	Sequence	18,361	50000	Nadir	Regional map	Coord.+Heading	Localization
AerialVL [ 6 ]	Single/Sequence	18.4K	300	Nadir	Regional map	Coord.	Retrieval/Localization

TABLE I: Comparison of UAV geo-localization benchmarks and released evaluation data. Sat. Extent reports approximate ground side lengths of satellite references in meters; regional-map extents may vary by location. The AerialVL image count refers to its visual place recognition subset. GeoVINS denotes its released evaluation data. N/A denotes information not reported in the corresponding source.

Fig. 2: Overview of SatFix. The satellite image and multiple UAV views are jointly encoded by the VGGT- $\Omega$ backbone. Satellite-grid features query UAV visual features through cross-attention, after which a coordinate-aware decoder predicts a spatial distribution over continuous position hypotheses. The heading branch estimates a unit heading vector for each UAV view.

Retrieval-based methods
Method	Localization Error (m)		Localization Success (%)
Method	Mean	Median	@10 m	@25 m	@50 m	@100 m
University-1652 baseline [ 1 ]	230.72	221.13	0.09	0.65	2.71	11.88
Sample4Geo [ 3 ]	171.12	141.28	0.08	0.57	2.78	21.13
MFRGN [ 17 ]	170.47	139.99	0.09	0.52	2.73	21.57
DMNIL [ 18 ]	165.24	134.95	0.09	0.50	2.54	21.84

TABLE II: Single-view UAV–satellite metric localization on all 45,306 University-Metric test frames. All methods use the same location-level split and evaluation protocol; localization success is reported in percent.

Method	Views	Position (m)		Heading ( ∘ )		SR (%)
Method	Views	Mean	Med.	Mean	Med.	@10 m	@25 m	@50 m	@100 m
VGGT- $\Omega$ zero-shot	3	270.98	232.94	41.04	27.61	0.40	3.54	15.53	30.43
VGGT- $\Omega$ zero-shot	5	304.04	252.35	40.32	23.82	0.45	4.15	16.83	31.87
VGGT- $\Omega$ zero-shot	7	311.26	251.32	39.04	22.76	0.73	4.58	18.81	33.92
VGGT- $\Omega$ zero-shot	9	319.01	253.69	39.04	22.72	0.61	4.48	18.82	34.30
VGGT- $\Omega$ fine-tuned	3	83.00	66.77	38.23	31.82	6.16	23.32	40.52	63.37

TABLE III: Localization and heading estimation in the multi-view setting on University-Metric for

T\in\{3,5,7,9\}

. Localization success is reported in percent.

Configuration	XY	Joint H–O	Heading	Mean (m)	Median (m)	@10 m	@100 m
CoordConv only	$\checkmark$			83.30	45.56	15.97	65.87
Joint H–O only		$\checkmark$		84.81	49.20	16.45	65.40
Heading only			$\checkmark$	85.64	47.20	15.19	64.94
w/o CoordConv		$\checkmark$	$\checkmark$	87.15	50.15	16.57	64.15
w/o Joint H–O	$\checkmark$		$\checkmark$	82.80	47.85	15.97	65.81
w/o Heading	$\checkmark$	$\checkmark$		84.71	48.66	17.08	65.32

TABLE IV: Component ablation on all 45,306 University-Metric test frames with single-view inputs. XY denotes coordinate channels, and Joint H–O denotes the joint heatmap–offset objective.

Loss Weight			Position (m)		Heading ( ∘ )
$\lambda_{\rm head}$	$\lambda_{\rm rel}$	$\lambda_{\rm cham}$	Mean	Median	Median
0.15	0.25	0.10	85.95	48.85	21.45
0.20	0.25	0.10	85.07	48.52	21.04
0.25	0.25	0.10	82.33	45.66	20.81
0.30	0.25	0.10	88.48	51.14	22.57
0.35	0.25	0.10	84.53	49.27	21.63

TABLE V: Loss-weight ablation on all 45,306 University-Metric test frames with single-view inputs. The position-loss coefficient is fixed to one; one coefficient is varied at a time while the others remain at their default values.

Fig. 3: Qualitative comparison of multi-view UAV–satellite localization on University-Metric. From one to seven UAV views, ground-truth locations are shown in yellow, predictions from the VGGT- $\Omega$ zero-shot baseline in magenta, and SatFix predictions in green. The UAV views used in each case are shown below the corresponding satellite map. The reported mean error is averaged over the displayed localization points within each example and is used only to summarize the qualitative result.

Train Crop	Test Crop	Extent (m)	Mean (m)	Median (m)	SR@50 m (%)
1024	768	$\sim$ 473	63.67	36.94	62.57
1024	1024	$\sim$ 630	82.33	45.66	52.08
1024	2048	$\sim$ 1261	199.15	183.99	1.55
768	768	$\sim$ 473	99.94	59.75	43.69
2048	2048	$\sim$ 1261	97.66	85.24	15.79

TABLE VI: Effect of the satellite extent on University-Metric. Train Crop and Test Crop are the pixel sizes of the square satellite crop; Extent is the corresponding physical side length in meters. All crops are resized to

512\times 512

Explore similar work

ReLATE: Reliability-Guided Evidence Fusion for Robust UAV--Satellite cross-view Geo-Localization
Jul 28, 2026Haochen Jiang, Jialei Pan, Yuzhe Sun +4Image Corruption Robustness Cross-View Geo-Localization
School of Electronics and Information Engineering, Harbin Institute of Technology, Harbin 150001, China · National Key Laboratory of Radar Detection and Sensing, Nanjing Research Institute of Electronics Technology, Nanjing 210039, China · School of Electrical and Electronic Engineering, University of Manchester, Manchester M13 9PL, United Kingdom
OffNadirLoc: Benchmark and Framework for Challenging UAV-to-Satellite Geo-Localization under Large Off-Nadir Views
Jul 22, 2026Qian Qiao, Wenye Liu, Ting Liu +2Cross-View Geo-Localization Geospatial Representation Learning
School of Computer Science, Northwestern Polytechnical University
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

Jul 28, 2026, cs.CVJ/K move · Enter open · S save

ReLATE: Reliability-Guided Evidence Fusion for Robust UAV--Satellite cross-view Geo-Localization

Haochen Jiang, Jialei Pan, Yuzhe Sun, Zhe Dong+3

School of Electronics and Information Engineering, Harbin Institute of Technology, Harbin 150001, China · National Key Laboratory of Radar Detection and Sensing, Nanjing Research Institute of Electronics Technology, Nanjing 210039, China · School of Electrical and Electronic Engineering, University of Manchester, Manchester M13 9PL, United Kingdom

Image Corruption Robustness Cross-View Geo-Localization Remote Sensing Benchmarks UAV Remote Sensing Feature Fusion

Open paper AskPDF arXiv

Unmanned aerial vehicle (UAV)-satellite cross-view geo-localization matches UAV images against satellite imagery and has achieved impressive accuracy on clean (non-degraded) image benchmarks. In real-world flights, however, UAV observations are frequently affected by adverse weather, illumination changes, platform motion, sensor noise, and compression, while the robustness of existing methods under such degradations remains largely unexamined. In this paper, we present UAVSat-Deg, a large-scale robustness benchmark for degraded UAV-satellite geo-localization, comprising University-1652-Deg and SUES-200-Deg. UAVSat-Deg covers 27 corruption types, including 19 core and 8 compound corruptions, at three severity levels, supports bidirectional drone-to-satellite and satellite-to-drone retrieval as well as multi-height UAV acquisition, and contains more than 11.7 million pre-generated corrupted test images. Benchmarking representative methods under this protocol reveals substantial robustness gaps, particularly under severe and compound corruptions. To address this problem, we propose ReLATE, a Reliable Evidence Learning framework with Adaptive Token Evidence Regulation, which realizes reliability-adaptive feature fusion during descriptor construction. ReLATE estimates a structure-smoothed reliability field over visual tokens, aggregates trustworthy local evidence, and adaptively integrates it into query-derived representations; the regulated query representations are then combined with the CLS-token and GeM-pooled branches to form the final cross-view descriptor. Across both test sets and retrieval directions, ReLATE achieves the best average corrupted-test performance among the compared methods while maintaining competitive accuracy on clean images. The code and dataset will be available at https://github.com/JHC626/ReLATE.

Abstract

Explore similar work

ReLATE: Reliability-Guided Evidence Fusion for Robust UAV--Satellite cross-view Geo-Localization

OffNadirLoc: Benchmark and Framework for Challenging UAV-to-Satellite Geo-Localization under Large Off-Nadir Views

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