OrthoTrack — Continuous 6-DoF UAV Trajectory Estimation via Orthophoto-Anchored Tracking

MovingDrone DATASET

OrthoTrack RESULTS

OrthoTrack is a training-free system that estimates continuous, absolute, metrically scaled 6-DoF UAV trajectories using only publicly available orthophotos and surface models as a map prior.

MovingDrone is a large-scale benchmark pairing photorealistic UAV sequences with dense 6-DoF ground truth and co-registered multi-modal geodata.

Abstract

Continuous 6-DoF pose estimation is essential for autonomous UAV operations. Yet, existing visual odometry and SLAM methods accumulate drift and yield only relative, up-to-scale trajectories. Single-frame geo-localization, in turn, discards temporal continuity and remains too slow for real-time use. We present OrthoTrack, a training-free system that estimates continuous 6-DoF UAV trajectories using only publicly available orthophotos and surface models as a map prior. OrthoTrack matches keyframes against the orthophoto and lifts correspondences to metric 3D via the surface model. It then propagates these map-anchored correspondences to intermediate frames with optical flow, producing absolute, metrically scaled poses at every frame without GPS or post-hoc alignment. We also introduce the MovingDrone Dataset, a large-scale benchmark pairing photorealistic UAV sequences with dense 6-DoF ground truth and co-registered multi-modal geodata including multi-temporal orthophotos. On MovingDrone and real-world benchmarks, OrthoTrack runs in real time on a single GPU. It outperforms all baselines by a large margin, even those receiving oracle scale and alignment. By relying on publicly available geodata, OrthoTrack enables deployment to new regions without site-specific adaptation.

How It Works

Click each step to see real outputs from the selected sequence.

First Frame Localization

First-Frame Initialization

The UAV image is matched against the full 1km×1km DOP tile using a coarse matcher.

Keyframe Localization

Three overlapping DOP crops are matched with RoMaV2-precise at full resolution. Correspondences from all crops are merged and deduplicated, then solved via PnP-RANSAC.

Inter-Frame Tracking

LK optical flow tracks 2D–3D correspondences frame-to-frame. Lifted DSM points are propagated via optical flow, and every frame gets an absolute 6-DoF metric pose via PnP on the propagated 2D↔3D pairs.

MovingDrone Dataset

Large-scale photorealistic 6-DoF UAV pose benchmark across Berlin with co-registered multi-modal geodata.

194

Sequences

332K

Frames

30

FPS

6-DoF

Ground Truth

RTK

GPS

20cm

DOP GSD

20cm

DSM GSD

20M

LiDAR PTS

15

DOP Years

LoD1/2

Buildings

Qualitative Results

All synchronized modalities in one workspace. Scrub the global timeline to update UAV, geodata 2D trajectory, and geodata 3D simultaneously.

UAV

RGB

Depth

Normals

1×

DOP Years

Loading…

1×

GeoData 2D

DOP

DSM

1×

GeoData 3D

Quantitative Results

0.67m

MovingDrone ATE

1.51m

UAVScenes ATE

1.26m

UAVD4L ATE

23.8 fps

Throughput

Pose Estimation Results

10 test sequences with independently surveyed DOP/DSM geodata. Oracle-aligned VO/SLAM and FM methods are shown alongside absolute geodata-based methods to make the alignment advantage explicit.

Method	Type	Align.	ATE↓ m	TE↓ m	RE↓ °	R@1↑ %	R@2↑ %	FPS↑
OrthoTrack (Ours) This work	Map	—	0.67	0.33	0.06	90.9	97.9	23.8
OrthoTrack loc. only	Map	—	4.83	0.30	0.06	95.8	99.2	1.9
OrthoLoC Dhaouadi et al., 2025	Map+Prior	—	19.78	0.50	0.08	85.4	95.4	0.5
LoD-Loc (retrained) Zhu et al., 2024	Map+Prior	—	10.04	5.25	1.10	10.8	30.3	3.3
DROID-SLAM Teed and Deng, 2021	VO/SLAM	pw	0.63	0.35	0.08	93.7	99.1	4.7
DROID-SLAM Teed and Deng, 2021	VO/SLAM	Sim(3)	0.34	0.22	0.47	88.8	89.7	4.7
DPVO Teed et al., 2024	VO/SLAM	Sim(3)	2.43	1.04	0.68	72.7	84.1	4.5
DA3-Nested Lin et al., 2025	FM	Sim(3)	2.74	1.83	2.57	28.7	51.4	11.3
VGGT-SLAM v2 Wang et al., 2025	FM	Sim(3)	11.94	9.28	15.88	0.0	0.0	3.1

Selected rows from the main comparison table. Sim(3) and pw rows use alignment during evaluation, whereas OrthoTrack reports absolute metric poses directly.

20 real-world sequences across Armenia and Hong Kong. Orthophotos and DSMs are rendered from the released LiDAR mesh, so the map-query domain gap is smaller than on MovingDrone.

Method	Type	Align.	ATE↓ m	TE↓ m	RE↓ °	R@1↑ %	R@2↑ %	FPS↑
OrthoTrack (Ours) This work	Map	—	1.51	1.38	0.49	20.6	88.0	6.7
OrthoTrack loc. only	Map	—	1.89	1.28	0.43	20.5	91.7	1.9
OrthoLoC Dhaouadi et al., 2025	Map+Prior	—	38.24	1.26	0.43	23.5	93.8	0.5
DROID-SLAM Teed and Deng, 2021	VO/SLAM	pw	8.03	2.53	0.80	5.4	32.1	9.1
ORB-SLAM3 Campos et al., 2021	VO/SLAM	Sim(3)	36.90	20.58	14.61	1.3	6.3	14.0
DPVO Teed et al., 2024	VO/SLAM	Sim(3)	100.92	85.78	53.88	0.1	0.9	9.2
DA3-Nested Lin et al., 2025	FM	Sim(3)	58.27	38.74	16.47	0.2	1.1	6.9
DUSt3R Wang et al., 2024	FM	Sim(3)	84.14	63.37	37.84	0.0	0.0	0.2
Pi3 Wang et al., 2025	FM	Sim(3)	102.80	80.39	53.75	0.0	0.0	17.3

UAVScenes uses LiDAR-mesh-derived geodata. The OrthoLoC row comes from the main paper; the remaining baseline rows follow the dedicated UAVScenes supplementary table.

19 sequences over Changsha, China, at 0.5 fps. The benchmark only provides a textured mesh, so orthophotos and DSMs are rendered from that mesh and the domain gap is therefore smaller than in MovingDrone.

Method	Type	Align.	ATE↓ m	TE↓ m	RE↓ °	R@1↑ %	R@2↑ %	R@5↑ %
OrthoTrack (Ours) This work	Map	—	1.26	1.01	0.26	50.7	96.5	99.8
OrthoLoC Dhaouadi et al., 2025	Map+Prior	—	1958.38*	1.09	0.34	43.4	84.8	88.0
DROID-SLAM Teed and Deng, 2021	VO/SLAM	Sim(3)	32.47	24.70	19.09	62.7	67.9	68.5
DPVO Teed et al., 2024	VO/SLAM	Sim(3)	33.74	26.24	22.40	57.0	66.1	68.4
Pi3 Wang et al., 2025	FM	Sim(3)	5.67	4.43	6.32	41.8	61.2	73.5
DUSt3R Wang et al., 2024	FM	Sim(3)	16.31	11.92	9.37	22.3	43.4	66.1
DA3-Nested Lin et al., 2025	FM	Sim(3)	36.81	31.72	13.41	49.6	64.5	76.2
ORB-SLAM3 Campos et al., 2021	VO/SLAM	Sim(3)	68.84	59.03	27.23	13.6	22.2	34.3

UAVD4L supplementary table. This benchmark reports R@5 instead of runtime, so the final column differs from the other two datasets.

Ablation Study

The main paper ablations are augmented here with supplementary analyses on sensor priors and map year choice.

Flow	ATE↓	TE↓	RE↓	R@2↑	FPS↑
LK Lucas and Kanade, 1981	0.67	0.33	0.06	97.9	23.8
RAFT Teed and Deng, 2020	0.66	0.37	0.08	98.4	10.0
SEA-RAFT Wang et al., 2024	1.93	0.37	0.08	97.9	14.1
NeuFlow v2 Zhang et al., 2025	0.82	0.33	0.06	97.9	17.6

Matcher	ATE↓	TE↓	RE↓	R@1↑	R@2↑	FPS↑
RoMa-v2 (Precise) Edstedt et al., 2025	0.67	0.33	0.06	90.9	97.9	23.8
RoMa-v2 (Base) Edstedt et al., 2025	0.90	0.49	0.10	86.4	95.9	26.3
GIM(DKM) Wang et al., 2024	2.56	0.52	0.10	89.1	97.9	21.7
L2M Liang et al., 2025	6.23	0.29	0.06	94.9	96.4	15.0
MASt3R Leroy et al., 2024	5.74	2.10	0.46	20.4	56.3	5.5

Input FPS	Subsample	ATE↓	TE↓	R@2↑	#KF	System FPS↑
30 fps	1×	0.67	0.33	97.9	15	23.8
10 fps	3×	0.58	0.39	97.9	11	19.2
3 fps	10×	0.68	0.42	97.8	8	11.1
2 fps	15×	0.71	0.38	97.3	8	8.7

Trigger Strategy	ATE↓	TE↓	R@2↑	#KF	FPS↑
Adaptive (Ours)	0.67	0.33	97.9	15	23.8
Fixed reprojection (2 px)	0.79	0.46	94.9	9	21.4
Fixed interval (K = 50)	0.53	0.35	99.3	28	19.0
Fixed interval (K = 150)	0.82	0.41	96.9	10	24.2
Point-count only	1.69	0.93	80.8	4	22.5

Sensor prior	ATE↓	TE↓	RE↓	R@1↑	R@2↑
GPS noise σh = 0 m	0.72	0.34	0.06	91.1	99.2
GPS noise σh = 3 m	0.63	0.33	0.07	93.8	98.8
GPS noise σh = 20 m	0.67	0.33	0.07	93.0	98.8
Yaw noise σyaw = 0°	0.66	0.34	0.06	92.4	98.8
Yaw noise σyaw = 30°	0.72	0.37	0.06	89.1	99.3
No prior	0.67	0.33	0.06	90.9	97.9

DOP year	TrueDOP	ATE↓	TE↓	RE↓	R@1↑	R@2↑
2025	Yes	0.67	0.33	0.06	90.9	97.9
2023	Yes	3.34	0.44	0.08	88.5	97.0
2020	Yes	0.93	0.65	0.15	78.4	94.5
2016	No	70.55	26.28	0.78	14.6	27.9
2013	Yes	10.22	2.28	0.67	56.6	78.8
2011	No	110.4	102.9	2.94	9.0	19.8

BibTeX

@misc{orthotrack2025,
  title     = {OrthoTrack: Continuous 6-DoF UAV Trajectory Estimation
               via Orthophoto-Anchored Tracking},
  author    = {Dhaouadi, Oussema and Bauer, Zuria and Meier, Johannes and Wysocki, Olaf and Pollefeys, Marc and Cremers, Daniel},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year      = {2026}
}