OrthoLoC: UAV 6-DoF Geodata Localization

• OrthoLoC is a geodata-centric framework that enables UAV 6-DoF localization and calibration by leveraging digital orthophotos (DOP) and digital surface models (DSM).
• It introduces Adaptive Homography Preconditioning (AdHoP) to refine cross-view matching between oblique UAV imagery and nadir orthophotos, significantly reducing matching and translation errors.
• The framework decouples pose estimation from image retrieval using a paired dataset across diverse scenes, allowing detailed evaluation of dense, semi-dense, and sparse matching techniques.

Searching arXiv for the OrthoLoC paper and closely related works to ground the article. OrthoLoC is a geodata-centric framework for UAV visual localization and calibration that replaces large posed-image databases and heavy 3D models with orthographic reference data, specifically digital orthophotos (DOP) and digital surface models (DSM). It was introduced as the first large-scale dataset specifically designed for 6-DoF localization and calibration using orthographic geodata, with a paired benchmarking protocol that decouples image retrieval from pose estimation and a refinement method, Adaptive Homography Preconditioning (AdHoP), for reducing the cross-view disparity between oblique UAV imagery and nadir orthophotos (Dhaouadi et al., 22 Sep 2025).

1. Definition and operating regime

OrthoLoC addresses accurate 6-DoF visual localization for UAVs in settings characterized by limited resources, including the absence of GNSS, constrained compute and storage, and often no internet connectivity. The central claim of the paradigm is that orthographic geodata constitute a lightweight alternative to two established baselines: retrieving the nearest pose from a large database of geo-tagged images, and localizing against heavy 3D models such as SfM or MVS reconstructions, LoD maps, or meshes. Compared to 3D meshes, geodata covering $0.265\ \text{km}^2$ require on the order of $30\times$ less memory than an $\sim 8\ \text{GB}$ mesh model, scale naturally to large areas, and do not require internet retrieval at runtime (Dhaouadi et al., 22 Sep 2025).

The method is formulated as full $SE(3)$ pose recovery from a perspective UAV image against orthographic references. In OrthoLoC, the UAV query image is matched to a paired orthophoto crop; the matched orthophoto coordinates are then lifted to 3D using DSM heights, and pose is solved from resulting 2D–3D correspondences. This design deliberately removes retrieval from the benchmark so that localization and calibration can be evaluated in isolation under controlled covisibility and controlled spatial offsets (Dhaouadi et al., 22 Sep 2025).

A common misconception is to treat OrthoLoC as a retrieval benchmark. It is not: retrieval is explicitly excluded from the evaluation protocol, and paired crops simulate retrieval imprecision via random offsets of $0\text{–}10\ \text{m}$ while preserving control over the visible region. Another misconception is that the framework is restricted to nadir UAV views. The dataset includes obliqueness angles with mean $14.6^\circ$, minimum $0^\circ$, and maximum $86.8^\circ$, covering both nadir and oblique imagery (Dhaouadi et al., 22 Sep 2025).

2. Dataset construction and benchmark organization

The dataset contains 16,427 paired samples, comprising approximately 16,425 UAV images, collected across 47 regions in 19 cities spanning Germany and the United States. Each sample pairs a query UAV image with georeferenced DOP and DSM rasters, and also includes point maps, visible 3D keypoints, and local meshes. The scenes cover urban, suburban, rural, and highway environments, and an augmented cross-domain variant introduces DOP and DSM from external governmental sources in order to induce appearance and structural domain gaps (Dhaouadi et al., 22 Sep 2025).

The ground truth is built from GPS/RTK and manually annotated GCPs with bundle-adjustment constraints. Reported GCP residual RMSE values are $0.023\ \text{m}$ in $x$, $30\times$0 in $30\times$1, $30\times$2 in $30\times$3, and $30\times$4 overall in 3D. All geometry is transformed into a local coordinate system to preserve privacy while maintaining exact relative geometry (Dhaouadi et al., 22 Sep 2025).

Component	Specification
Modalities per sample	UAV RGB image $30\times$5, orthophoto raster $30\times$6, elevation raster $30\times$7, point map $30\times$8, visible 3D keypoints $30\times$9, local mesh $\sim 8\ \text{GB}$0
Standard raster size	$\sim 8\ \text{GB}$1 for DOP/DSM
Query image size	$\sim 8\ \text{GB}$2 or $\sim 8\ \text{GB}$3, undistorted
Spatial resolution	Mean $\sim 8\ \text{GB}$4, min $\sim 8\ \text{GB}$5, max $\sim 8\ \text{GB}$6
Splits	Train 13,347; Val 1,489; Test In-Place 176; Test Out-Place 1,415
Domain types	same-domain 10,923; DOP cross-domain 4,698; DOP+DSM cross-domain 806

The paired structure is one of the dataset’s defining design decisions. Because the orthographic crop is known and covisibility is very high by construction—minimum $\sim 8\ \text{GB}$7, near $\sim 8\ \text{GB}$8 in the standard paired crops—the benchmark can isolate feature matching and pose recovery behavior without conflating them with errors from large-scale retrieval. This also enables systematic ablations over covisibility, spatial offset, raster resolution, and domain shift (Dhaouadi et al., 22 Sep 2025).

3. Geometric formulation and localization pipeline

The OrthoLoC pipeline models localization as perspective-to-orthographic matching followed by 2D–3D pose recovery. The UAV camera pose is written as $\sim 8\ \text{GB}$9, with $SE(3)$0 and $SE(3)$1, and projection is modeled as

$SE(3)$2

where images are undistorted and lens distortion is ignored in pose estimation. The orthographic reference is represented by a nadir camera induced by the raster grid with origin $SE(3)$3 and scales $SE(3)$4, so pixel coordinates map to metric coordinates by $SE(3)$5, $SE(3)$6 (Dhaouadi et al., 22 Sep 2025).

The practical pipeline is backbone-agnostic. It first matches features between the UAV image and the orthophoto. Orthophoto correspondences are then lifted to 3D with DSM heights, producing 3D–2D pairs. Pose is initialized with RANSAC-EPnP using a $SE(3)$7 inlier threshold. For calibration, Levenberg–Marquardt refinement jointly optimizes extrinsics and intrinsics by minimizing a robust reprojection objective,

$SE(3)$8

with Huber loss $SE(3)$9. For localization-only tasks, $0\text{–}10\ \text{m}$0 is held fixed and only $0\text{–}10\ \text{m}$1 is optimized (Dhaouadi et al., 22 Sep 2025).

A planar model also appears in the method through a plane-induced homography,

$0\text{–}10\ \text{m}$2

which reduces perspective disparity between the oblique UAV image and the orthophoto under a local planar approximation. This is not the final pose solver; beyond a single plane, elevation is restored through DSM lifting, and full $0\text{–}10\ \text{m}$3 is recovered through PnP with RANSAC and LM refinement (Dhaouadi et al., 22 Sep 2025).

OrthoLoC evaluates dense, semi-dense, and sparse matchers. The reported matcher families are GIM+DKM, RoMa, DUSt3R, and MASt3R in the dense category; LoFTR, eLoFTR, and XoFTR in the semi-dense category; and SuperPoint+SuperGlue, SuperPoint+LightGlue, DeDoDe, and XFeat in the sparse category. Matching verification employs confidence thresholds of at least $0\text{–}10\ \text{m}$4, field-of-view filtering, and LO-RANSAC during PnP (Dhaouadi et al., 22 Sep 2025).

4. AdHoP and the calibration problem

AdHoP, or Adaptive Homography Preconditioning, is the principal algorithmic addition of OrthoLoC. It is method-agnostic and training-free. Starting from an initial set of image-to-orthophoto matches $0\text{–}10\ \text{m}$5, it estimates a homography with normalized DLT and RANSAC, warps the orthophoto, repeats feature matching in the warped domain, maps the improved matches back with $0\text{–}10\ \text{m}$6, lifts them to 3D via the DSM, and resolves pose with RANSAC-EPnP followed, optionally, by LM refinement. The refinement is accepted only if the mean reprojection error decreases (Dhaouadi et al., 22 Sep 2025).

The purpose of AdHoP is to reduce the perspective disparity between orthographic and oblique views before the second matching round. The paper attributes to it improvements of up to $0\text{–}10\ \text{m}$7 in matching error, up to $0\text{–}10\ \text{m}$8 in translation error, and substantial recall gains under specific configurations and datasets. Selected examples include a $0\text{–}10\ \text{m}$9 matching error decrease for XFeat*, a $14.6^\circ$0 translation error reduction for DKM, and a $14.6^\circ$1 absolute increase in $14.6^\circ$2 recall for RoMa (Dhaouadi et al., 22 Sep 2025).

The calibration part of OrthoLoC is explicitly constrained by a focal length–translation ambiguity. In approximately planar aerial scenes, the depth term $14.6^\circ$3 dominates, and scaling focal length $14.6^\circ$4 and translation $14.6^\circ$5 by the same factor leaves the perspective projection effectively unchanged. The paper therefore argues that calibration is ill-posed unless correspondences exhibit strong depth variation and spatial diversity. AdHoP helps not by eliminating the ambiguity analytically, but by spreading matches across the image and across differing depths, which weakens the degeneracy in practice (Dhaouadi et al., 22 Sep 2025).

This point is operationally important. OrthoLoC is not only a localization benchmark; it is also a calibration benchmark, but the paper treats calibration as fundamentally more difficult than pose estimation. The reported relative focal length error is approximately $14.6^\circ$6 for the best configuration, achieved together with translation error of approximately $14.6^\circ$7, while calibration recall remains modest because of the focal–$14.6^\circ$8 ambiguity (Dhaouadi et al., 22 Sep 2025).

5. Empirical findings, dominant error sources, and practical limits

The strongest reported localization results are obtained by GIM+DKM with AdHoP. On challenging test sets, this configuration achieves translation error of approximately $14.6^\circ$9, rotation error of approximately $0^\circ$0, and $0^\circ$1 recall of approximately $0^\circ$2. Sparse baselines such as SuperPoint+SuperGlue and SuperPoint+LightGlue achieve approximately $0^\circ$3 translation error and approximately $0^\circ$4 rotation error, but with lower recall than dense methods. Semi-dense matchers such as LoFTR and XoFTR underperform in aerial cross-domain settings, with recall reported as at most approximately $0^\circ$5 (Dhaouadi et al., 22 Sep 2025).

The paper identifies three variables as especially influential: domain shift, resolution, and covisibility. Domain shift is explicitly decomposed into visual shifts in DOP—color, illumination, shadows, seasonal change, and temporal offset—and structural shifts in DSM due to construction, demolition, vegetation growth, and infrastructure change. Using GIM+DKM with AdHoP, translation error increases by approximately $0^\circ$6 under DOP cross-domain conditions and by approximately $0^\circ$7 when both DOP and DSM are cross-domain; the example given is $0^\circ$8 (Dhaouadi et al., 22 Sep 2025).

Resolution ablations show that the framework remains robust down to approximately $0^\circ$9 rasters, whereas at $86.8^\circ$0 translation error increases by approximately $86.8^\circ$1 and rotation error by approximately $86.8^\circ$2. Query resolution exhibits a similar pattern, so moderate downscaling offers computational savings with limited accuracy loss. Runtime and memory analyses support the low-resource framing: dense matchers plus AdHoP typically add approximately $86.8^\circ$3 per query, for example GIM+DKM increasing from $86.8^\circ$4 to $86.8^\circ$5, while remaining far smaller in memory footprint than 3D meshes (Dhaouadi et al., 22 Sep 2025).

Covisibility is the strictest geometric constraint. The paired benchmark is designed with very high overlap, but ablations show sharp performance degradation below approximately $86.8^\circ$6 overlap because the correspondence set becomes too poorly distributed for well-conditioned PnP. The paper emphasizes that this is not merely a matter of fewer matches; it is specifically the poor spatial spread of correspondences that makes pose estimation underconstrained (Dhaouadi et al., 22 Sep 2025).

A further misconception is that 2.5D orthographic geodata are inherently insufficient for precise localization. OrthoLoC argues against this in a restricted sense: on same-domain data, ground-truth geometry-aware filtering shows that 2.5D geodata are sufficient for perfect localization when ambiguous vertical-structure correspondences are excluded. This suggests that the limiting factors are not only map dimensionality, but also cross-view matching ambiguity and geometric inconsistency induced by vertical structures and domain shift (Dhaouadi et al., 22 Sep 2025).

6. Terminological scope and related uses

In its strict and primary sense, OrthoLoC denotes the UAV dataset, protocol, and geodata-based localization paradigm introduced in "OrthoLoC: UAV 6-DoF Localization and Calibration Using Orthographic Geodata" (Dhaouadi et al., 22 Sep 2025). The name, however, appears in additional and distinct contexts in later literature syntheses.

In orthopedic imaging papers, "OrthoLoC" is used as an overview label for orthopedic localization or "Location-of-Contact" capabilities rather than as the title of a formally introduced standalone method. In "Multi-task Localization and Segmentation for X-ray Guided Planning in Knee Surgery," the term is used to frame orthopedic landmark localization, bone segmentation, and planning automation around expert-level femoral drill-site planning on lateral X-rays (Kordon et al., 2019). In "4D Virtual Imaging Platform for Dynamic Joint Assessment via Uni-Plane X-ray and 2D-3D Registration," it denotes dynamic localization of joint pose and tibiofemoral contact under physiological load, including medial–lateral contact variance and distance-difference measures in post-TKA assessment (Tang et al., 22 Aug 2025). In the synthesis of "DualVision ArthroNav," the same label is applied to an arthroscopy navigation system combining a stereo external camera with a monocular arthroscope to resolve scale ambiguity and drift in image-based intra-articular localization and reconstruction (Shu et al., 12 Nov 2025). A related synthesis of "DeepLOC" uses OrthoLoC as a broader orthopedic localization-and-classification concept for pathology detection in wrist X-rays (Dibo et al., 2023).

A separate disambiguation concerns the PEFT method "LoCO: Low-rank Compositional Rotation Fine-tuning." That paper states that the query term "OrthoLoC" can be taken as an informal alias emphasizing the orthogonality of LoCO, but it also states that there is no separate method or naming distinction in the paper itself, which consistently uses "LoCO" (Nguyen et al., 15 May 2026).

These secondary usages do not alter the primary meaning of OrthoLoC in the arXiv literature. They instead show that the name has become attractive wherever orthographic, orthopedic, or orthogonality-centered localization is being discussed. The formally introduced entity remains the UAV 6-DoF localization and calibration framework based on orthographic geodata, paired benchmarking, and AdHoP refinement (Dhaouadi et al., 22 Sep 2025).