PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Jiarui Zeng, Kun Shi, Chiman Vong, Zhedong Zheng

arXiv:2610.11049v1cs.CV

TL;DR

Gallery-based cross-view localization returns discrete tile indices rather than continuous position and heading, while metric UAV localization within a provided satellite region remains a target. SatFix predicts map-frame position and heading from one or more UAV views. On University-Metric, its nine-view results include 21.96 m median position error and 8.73° median heading error.

  • Problem

    Gallery-based cross-view localization depends on gallery sampling and returns a tile index rather than continuous position or heading.

    Source: §I · p. 1

  • Method

    SatFix grounds geometric features in the satellite coordinate frame and estimates continuous position and heading from one or more UAV views.

    Source: B. Position and Heading Modeling · p. 4, §VI · pp. 7–8

  • Results

    With nine views, SatFix reaches 21.96 m median position error and 8.73° median heading error on University-Metric.

    Source: §VI · p. 8

  • Takeaways & Limitations

    SatFix achieves lower position and heading errors than zero-shot and fine-tuned VGGT-Ω baselines across the reported multi-view settings.

    Source: B. Ablation Studies and Further… · p. 6, §VI · p. 8

    The evaluation assumes satellite coverage and a coarse prior bounding the search to the provided crop; robustness to map changes and incomplete maps remains untested.

    Source: B. Ablation Studies and Further… · p. 7

Abstract

from arXiv · show

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-$\Omega$. 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-$\Omega$ baseline, SatFix reduces median position error by 34.0% and nine-view median heading error from $25.43^\circ$ to $8.73^\circ$.

I. INTRODUCTION

The paper targets continuous position and heading estimation from oblique UAV views within a geo-referenced satellite region, addressing gaps in both benchmark coverage and efficient feed-forward methods. It introduces University-Metric and SatFix, reporting single- and multi-view localization results.

Motivation

  • Existing benchmarks often use gallery retrieval over satellite tiles about 118–120 m per side or near-nadir UAV views, while this work studies regional localization from oblique views.Retrieval yields a tile index rather than continuous position or heading.Source: §I · p. 1

Region-scale oblique benchmark

  • University-Metric pairs oblique UAV imagery with regional satellite maps up to 10.7× longer per side than University-1652 tiles, with per-frame position and heading labels.Its location-level split includes 631 training, 70 validation, and 839 test locations.Source: §I · pp. 1–2

Grounding by cross-attention

  • SatFix uses satellite-grid features to query UAV features, then decodes continuous position and heading without scene reconstruction, tile retrieval, a 3D map, auxiliary sensors, or test-time alignment.A single model supports single-view and multi-view inputs.Source: §I · pp. 1–2

Accuracy and efficiency

  • With nine views, SatFix reaches 21.96 m median position error and 8.73° median heading error; single-view inference takes under 0.1 s per query.For one view, 52.08% of frames are within 50 m.Source: §I · p. 2

II. RELATED WORK

Prior UAV cross-view localization largely retrieves a gallery item, while metric approaches increasingly predict geographic positions using maps or visual geometry. SatFix instead predicts continuous map-frame position and heading directly from a provided satellite region and oblique UAV views.

Retrieval-based Cross-view Geo-localization

  • Retrieval-based geo-localization returns a gallery item, so its geographic estimate depends on gallery sampling density.Methods have advanced through compressed map representations, multimodal fusion, local feature aggregation, temporal consistency, and semantic mapping.Source: §II · p. 2

Metric UAV Geo-localization

  • Metric localization includes map-based systems and methods that estimate position and heading, but several approaches rely on reconstruction, retrieval, or other map representations.Bearing-UAV jointly estimates position and heading; AnyVisLoc studies low-altitude multi-view localization.Source: §II · p. 2
  • SatFix predicts continuous map-frame (x, y, θ) from a geo-referenced RGB satellite region and one or a few oblique UAV views, without dense 3D maps or long-horizon tracking.Its geometric features are grounded directly in the satellite coordinate frame rather than through explicit scene reconstruction and candidate-map retrieval.Source: §II · p. 2

III. UNIVERSITY-METRIC: DATASET AND TASK SETTING

University-Metric supports metric localization from one or more UAV views by pairing frame-level position and heading labels with geo-referenced regional satellite crops. The task assumes a coarse geographic prior that bounds the search to the provided region.

Benchmark setting and task

  • University-Metric provides continuous position and heading labels for every UAV frame, enabling metric localization from a single view or a short clip.Its default satellite crop spans 630 m per side, or 28× the ground area of the original per-building tiles.Source: §III · p. 2
  • The benchmark is built from University-1652 UAV tours and trajectory/KML metadata, with regional satellite references re-collected over larger extents.The largest crops cover approximately 114× the ground area of the original tiles.Source: §III · p. 2

Task formulation

  • The task predicts a continuous coordinate within the provided geo-referenced region rather than retrieving a discrete gallery item; “absolute” refers to the satellite coordinate frame.The operating region is specified by a coarse geographic prior.Source: §III · pp. 2–3

Annotations and coordinate representation

  • The benchmark splits source locations into 631 training, 70 validation, and 839 test locations, with no tour or source location shared across splits.Each location’s satellite crop is centered at its geo-referenced map center and shared across its 54 UAV frames.Source: §III · p. 3

Evaluation protocol and satellite extent

  • Position errors are converted to meters using location-specific satellite-map scale, while heading error uses the minimum circular angular difference.The default crop covers about 630 m per side, and satellite and UAV images are resized to 512 × 512.Source: §III · p. 3

IV. METHOD

SatFix matches UAV visual features to the satellite grid, then decodes the matched features into continuous position and heading estimates.

Method

  • SatFix first matches UAV visual features to the satellite grid, then decodes the matched features into continuous positions and headings.Source: §IV · p. 3

A. Geometry-grounded Localization

SatFix jointly encodes satellite and UAV imagery, using satellite-grid features to gather UAV evidence through cross-attention. A coordinate-aware decoder predicts continuous position hypotheses, while a separate branch estimates each view’s heading.

A. Geometry-grounded Localization

  • Satellite-grid features query UAV features through cross-attention, and the fused representation retains the map-grid layout for localization.The system jointly encodes the satellite image and UAV views, then fuses matched features with satellite features and camera conditioning.Source: A. Geometry-grounded Localization · p. 3, Figure 2 · p. 4
  • The heading branch estimates a unit heading vector for each UAV view alongside the continuous position hypotheses.Source: Figure 2 · p. 4
  • A coordinate-aware decoder adds explicit map coordinates to grid-cell features before predicting a spatial probability map and offsets.The probability map is normalized so its entries sum to one.Source: A. Geometry-grounded Localization · p. 4

B. Position and Heading Modeling

SatFix forms continuous position hypotheses from grid cells and predicted offsets, using a probability map for inference and differentiable supervision. Heading is estimated as a unit vector, and multi-view training adds trajectory constraints.

B. Position and Heading Modeling

  • At inference, SatFix selects the highest-probability cell hypothesis and maps it to the satellite crop; training also uses the probability-weighted expected position.Each hypothesis adds a predicted offset to a cell center, yielding a continuous position.Source: B. Position and Heading Modeling · p. 4
  • Heading uses the position probability map to pool localization features, then predicts a normalized two-dimensional vector whose bearing is recovered with atan2.The vector representation avoids the discontinuity between 0° and 360°.Source: B. Position and Heading Modeling · p. 4
  • The position objective combines joint, heatmap, offset, and expected-position losses, with auxiliary terms stabilizing position learning.The joint distance term weights a cell strongly only when its offset is accurate, encouraging sub-cell precision at high-scoring locations.Source: B. Position and Heading Modeling · pp. 4–5
  • For multiple views, the relative-displacement loss compares predicted and ground-truth displacement between consecutive views using normalized positions.The trajectory terms apply only when T > 1.Source: B. Position and Heading Modeling · pp. 4–5
  • A symmetric Chamfer term constrains the overall spatial layout by comparing predicted and ground-truth point sets without temporal ordering.Its mean and nearest-neighbor searches cover all views in the clip.Source: B. Position and Heading Modeling · pp. 4–5

V. EXPERIMENT

Experiments use University-Metric with a location-level train, validation, and test split and a default 630 m satellite crop. The evaluation compares methods under a shared coordinate frame and protocol, with separate single- and multi-view settings.

V. EXPERIMENT

  • All experiments use University-Metric’s location-level split and, unless otherwise stated, a 630 m satellite reference crop.Training samples one, three, five, seven, or nine UAV views; multi-view test clips are sampled from the same 839 test locations in temporal order.Source: §V · p. 5
  • Table II covers single-view metric localization on 45,306 test frames, while Table III covers multi-view localization and heading for T ∈ {3, 5, 7, 9}.Both captions specify that localization success is reported in percent.Source: Table II · p. 5, Table III · p. 5
  • The comparison evaluates retrieval and direct-localization methods under the same split, satellite extent, coordinate transformation, and metric evaluator.Retrieval methods return the center of the top-ranked tile from 3,025 candidates for the default crop, while direct methods retain continuous position outputs.Source: §V · pp. 5–6

A. Comparison with the State-of-the-Art Methods

On single-view University-Metric evaluation, SatFix lowers median position error relative to fine-tuned VGGT-Ω. The comparison also reports substantially larger errors for zero-shot VGGT-Ω and AnyVisLoc.

A. Comparison with the State-of-the-Art Methods

  • 34.0% lower median position error: SatFix records 45.66 m versus 69.17 m for fine-tuned VGGT-Ω, and lowers mean error from 88.15 m to 82.33 m.Across all 45,306 test frames, SatFix has 17.34%, 34.88%, 52.08%, and 66.18% localization success at 10, 25, 50, and 100 m, respectively.Source: A. Comparison with the… · p. 6
  • AnyVisLoc has 164.15/137.04 m mean/median error versus SatFix’s 82.33/45.66 m, while zero-shot VGGT-Ω has 212.64/196.49 m mean/median error.Source: A. Comparison with the… · p. 6
  • Continuous outputs alone do not resolve the oblique cross-view differences, and the reported comparison attributes further improvement to SatFix’s localization modules beyond fine-tuning.Source: A. Comparison with the… · p. 6

B. Ablation Studies and Further Discussion

The ablations examine multi-view gains, model components, loss weights, and satellite-map extent. Results show improved localization with more views, while performance depends on training choices and the provided map region.

Multi-view setting

  • At nine views, SatFix reaches 21.96 m median position error and 8.73° median heading error; fine-tuned VGGT-Ω remains less accurate for T=3, 5, 7, and 9.From one to nine views, median position error falls from 41.85 m to 21.96 m and median heading error from 16.47° to 8.73°; gains diminish beyond five views.Source: B. Ablation Studies and Further… · p. 6

Effect of localization components

  • Adding the joint position objective reduces mean/median position error from 82.80/47.85 m to 82.33/45.66 m and raises SR@10 m from 15.97% to 17.34%.Across evaluated configurations, the full model has the lowest mean position error and highest localization success rates; coordinate channels alone have nearly identical median error, and removing heading supervision changes SR@10 m only slightly.Source: B. Ablation Studies and Further… · p. 6, Table IV · p. 6

Sensitivity to loss weights

  • Increasing λrel from 0.25 to 0.30 or 0.35 raises mean position error from 82.33 m to 93.53 or 93.21 m, while tested λcham values change it only moderately.Median heading error is 20.57° at λcham=0.08 and 20.81° at the default 0.10; small differences among nearby settings are treated as sensitivity trends, not precise rankings.Source: B. Ablation Studies and Further… · p. 6, Table V · p. 7

Effect of satellite extent

  • Reducing satellite extent from 630 m to 473 m lowers mean error from 82.33 m to 63.67 m, while increasing it to about 1261 m raises mean error to 199.15 m and lowers SR@50 m to 1.55%.Training on 2048-pixel crops reduces mean error at the larger extent from 199.15 m to 97.66 m; matching 768-pixel crop training does not outperform training on 1024-pixel crops.Source: B. Ablation Studies and Further… · pp. 6–7, Table VI · p. 7

Qualitative comparison

  • Figure 3 compares ground-truth locations with zero-shot VGGT-Ω and SatFix predictions across examples using one to seven views, with per-example mean error as a qualitative summary.Ground truth is yellow, zero-shot predictions magenta, and SatFix predictions green; the corresponding UAV views appear below each satellite map.Source: Figure 3 · p. 7

Scope and limitations

  • The evaluation assumes satellite coverage and a coarse prior bounding the search to the provided crop; robustness to map aging, seasonal change, sensor shift, and incomplete maps remains untested.The authors also state that the experiments measure localization accuracy, not closed-loop navigation, odometry integration, or UAV–UGV cooperative routing.Source: B. Ablation Studies and Further… · p. 7

VI. CONCLUSION

SatFix addresses absolute metric UAV localization from oblique views within a geo-referenced satellite region, and University-Metric provides enlarged regional references with continuous labels. Across reported multi-view settings, SatFix outperforms the VGGT-Ω baselines, reaching 21.96 m median position error and 8.73° median heading error with nine views.

Benchmark

  • University-Metric pairs University-1652-derived UAV tours with enlarged regional satellite references, location-level train/validation/test splits, and continuous position and viewing-heading annotations.The benchmark is derived from University-1652.Source: §VI · pp. 7–8

Method

  • SatFix grounds feed-forward geometric representations in the satellite coordinate frame and jointly estimates continuous position and heading with one model for single- and multi-view inputs.Source: §VI · pp. 7–8

Results

  • Across reported multi-view settings, SatFix has lower position and heading errors than zero-shot and fine-tuned VGGT-Ω, reaching 56.12/21.96 m mean/median position error and 8.73° median heading error with nine views.Source: §VI · p. 8

Related papers

Loading 2610.11049v1…