MF-UAVPose6D: A Model-Free Monocular 6-DoF Pose Estimation Framework for Fixed-Wing UAVs

Abstract: For uncrewed aerial vehicles (UAVs), estimating six-degree-of-freedom (6-DoF) poses is essential for airspace situational awareness, target tracking, and counter-UAV operations. However, non-cooperative targets usually lack computer-aided design (CAD) models and keypoint priors, making existing model-based or keypoint-matching methods difficult to apply reliably. To address these challenges, this paper proposes MF-UAVPose6D, a model-free monocular 6-DoF pose estimation framework for fixed-wing UAVs. During inference, the method takes only a single red-green-blue (RGB) image and camera intrinsics as input. It first obtains a stable target anchor through heatmap-guided center localization, introduces a Perspective-Aware Module (PAM) to model observation-ray priors, exploits Dynamic Topological Sampling (DTS) to complement weak structural cues from the wings, fuselage, and tail, and adopts a decoupled translation-rotation pose decoding mechanism to estimate the 6-DoF pose. In addition, we construct the FW-UAV6DPose synthetic dataset, which covers fixed-wing UAV observations across diverse distances, viewpoints, and poses. Experimental results show that MF-UAVPose6D achieves accurate and efficient monocular 6-DoF pose estimation without requiring CAD models, and demonstrates strong robustness in long-range rotation estimation, depth recovery, and joint pose evaluation.

• The paper introduces MF-UAVPose6D, an end-to-end model-free framework that combines heatmap-based localization, perspective-aware conditioning, dynamic topological sampling, and algebraic depth decoding to estimate UAV pose from one RGB image and camera intrinsics.
• The method achieves 4.96° mean rotation error, 7.28 m mean translation error, 82.8% Pose@10°/5%, and 4.6 ms inference time on the synthetic FW-UAV6DPose benchmark without CAD models or external detection boxes.
• Ablation results show that perspective-aware conditioning and dynamic topological sampling substantially improve rotation accuracy, while algebraic depth decoding strengthens long-range depth estimation, although real-world transfer remains unvalidated.

Problem setting and motivation

Estimating the six-degree-of-freedom (6-DoF) pose of a fixed-wing UAV from a single RGB image is a perception task of direct relevance to airspace situational awareness, target tracking, and counter-UAV operations. The dominant paradigm for UAV pose estimation—detect the target, extract structural keypoints such as wingtips or fuselage endpoints, then solve with PnP—presupposes a known target model, known physical dimensions, or reliably visible keypoints. These assumptions fail for non-cooperative targets at long range, where the aircraft occupies roughly 1% of the image, texture is weak, and keypoint localization errors are amplified by the geometric solver. Generalizable model-free methods such as MegaPose, FoundationPose, SAM-6D, and Any6D remove the CAD dependency but were developed for close-range tabletop objects with large image footprints, and their transfer to hundred-meter-scale aerial observation is untested.

The paper under review addresses this gap with MF-UAVPose6D, an end-to-end monocular framework that takes only an RGB image and camera intrinsics at inference time. The work also contributes FW-UAV6DPose, a synthetic dataset built on AirSim/Unreal Engine covering 100–500 m observation distances.

Framework architecture

MF-UAVPose6D uses a RepViT backbone producing multi-scale features fused at stride 8. Four components form the pipeline:

• Heatmap-guided center localization: a lightweight head predicts a center probability map; a Gaussian-windowed soft-argmax converts the discrete peak into a continuous sub-pixel anchor $(u_c, v_c)$ and a pooled center feature $f_c$. This avoids the instability of hard argmax decoding on small targets.
• Perspective-Aware Module (PAM): the anchor is back-projected through $K^{-1}$ into a unit observation ray $\mathbf{d}$, which conditions the center embedding via channel-wise FiLM-style modulation ($$f_{\mathrm{pam}} = f_v \odot \mathrm{Sigmoid}(\gamma) + \beta$$). The motivation is that image-plane position induces viewing-angle variation that can be confounded with true 3D rotation.
• Dynamic Topological Sampling (DTS): from $f_{\mathrm{pam}}$, the network predicts 2D bounding-box scale via log-scale offsets, then bilinearly samples features at eight points (four edge midpoints, four corners) plus a $7\times7$ dense grid over the predicted extent. This injects wing/fuselage/tail contour cues without any CAD prior.
• Decoupled pose decoding: separate MLP branches regress translation and rotation, avoiding multi-task interference between scale-sensitive translation and structure-sensitive rotation.

Rotation is decoded as a 6D continuous representation followed by Gram–Schmidt orthogonalization into an allocentric rotation relative to the viewing ray; the egocentric camera-frame rotation is recovered by composing with a ray rotation matrix $$R_{\mathrm{ego}} = R_{\mathrm{ray}} \cdot R_{\mathrm{allo}}$$. This factorization explicitly removes the perspective coupling between image position and apparent orientation.

Algebraic perspective depth decoder

The most consequential design choice is the algebraic perspective depth decoder (APDD). Rather than regressing absolute depth $Z$ directly—an ill-posed mapping from subtle appearance changes to a large distance range—the network predicts a logarithmic implicit physical scale $S = \exp(s)$, learned purely from supervision without any CAD-derived dimension. Depth is then decoded geometrically as

$f_c$0

with lateral offsets refined within $f_c$1 box-widths around the anchor and back-projected through the intrinsics. Training uses a logarithmic depth loss plus depth-normalized planar Smooth L1, alongside penalty-reduced focal loss for the heatmap and SO(3) geodesic loss with a 6D regularization term ($f_c$2) for rotation.

Dataset: FW-UAV6DPose

The dataset contains 15,325 annotated images across 161 sequences (7,725 train / 2,400 val / 4,500 test), rendered at $f_c$3 in BOP format with masks, intrinsics, visible bounding boxes, and ground-truth 6-DoF poses. Target distances span approximately 100–500 m, and the average visible bounding-box area is only about 1.20% of the image. Flight dynamics drive the UAV attitude, and the virtual camera adaptively tracks the target. The dataset is released publicly. A caveat worth noting: all quantitative results are obtained in simulation, so no claim about real-world transfer is established by this paper alone.

Comparative results

On the FW-UAV6DPose test set, MF-UAVPose6D achieves a mean rotation error of 4.96°, mean translation error of 7.28 m, Pose@10°/5% of 82.8%, ADD-0.5d of 88.5%, and a runtime of 4.6 ms per frame:

Method	CAD required	$f_c$4	$f_c$5	Pose@10°/5%	Runtime
DronePose	Yes	29.48°	28.89 m	43.1%	2.3 ms
GDR-Net	Yes	9.15°	7.9 m	74.62%	6.8 ms
PoET	No	45.24°	25.29 m	24.3%	106.2 ms
SC6D-Pose	No	9.36°	3.26 m	73.7%	13.2 ms
MF-UAVPose6D	No	4.96°	7.28 m	82.8%	4.6 ms

Two observations qualify these numbers. First, SC6D-Pose attains better translation error (3.26 m vs. 7.28 m) and ADD-0.5d (99.7% vs. 88.5%), but requires an external 2D detection box at inference; MF-UAVPose6D's advantage lies in the fully self-contained input specification and joint success rate. Second, the comparison against CAD-based methods conflates prior availability with architecture, though the result still indicates that explicit models do not confer an advantage at these ranges.

Ablation evidence

Each module is ablated with all other settings held fixed, and the gains are consistent:

• PAM reduces mean rotation error from 6.91° to 4.96° and raises Rot<10° from 82.4% to 90.9%. Distance-wise analysis shows the benefit grows with range (e.g., −3.42° mean error in the 300–400 m band), supporting the claim that ray conditioning matters precisely when appearance cues degrade. Notably, PAM slightly lowers ADD-0.5d (89.5% → 88.5%), indicating its contribution is specific to rotation rather than overall alignment.
• DTS reduces mean rotation error from 7.79° to 4.96°, with the largest single gain at 200–300 m (−5.03°). Its effect on ADD-0.5d is inconsistent, confirming it contributes topological/rotational cues rather than translation accuracy.
• APDD versus direct depth regression lowers $f_c$6 from 7.67 m to 7.28 m and $f_c$7 from 7.43 m to 6.96 m, and lifts Pose@10°/5% from 79.4% to 82.8%. The contrast is sharpest at 400–500 m, where Direct-Z degrades to 16.58 m translation error and 62.6% Pose@10°/5%, while APDD retains an $f_c$8 of 15.59 m and improves ADD-0.5d by 4.8%.

Heatmap visualizations further show stable center responses concentrated on the airframe despite weak-texture sky backgrounds, which matters because every downstream geometric computation depends on the anchor.

Sequential consistency

On two continuous test sequences, frame-wise predictions track ground-truth trends closely. In the 100–300 m sequence, per-component RMSEs are 0.34/0.14/1.62 m for $f_c$9 and 2.36°/0.77°/2.55° for Roll/Pitch/Yaw. At 400–500 m, RMSEs grow to 1.60/0.51/5.26 m and 3.19°/5.90°/2.76° respectively—a quantified long-range degradation, particularly in pitch and depth, though trajectory overlap remains high. The paper does not apply temporal filtering or smoothing; all results are effectively per-frame, which strengthens the interpretation that the reported consistency reflects the estimator rather than sequence-level post-processing.

Limitations and open questions

Several limitations are acknowledged or evident. All training and evaluation data are synthetic; generalization to real imagery with sensor noise, motion blur, and atmospheric effects remains unverified. The implicit scale $K^{-1}$0 in APDD is dataset-specific—it is learned under the depth distribution of FW-UAV6DPose and clamped to $K^{-1}$1—so cross-dataset depth transferability is an open question. The evaluation covers a single target category (fixed-wing UAVs), leaving multi-class or rotary-wing extension unaddressed. Finally, the paper itself notes that future work targets longer ranges, where the observed degradation in depth and pitch RMSE suggests current limits.

Conclusion

MF-UAVPose6D demonstrates that model-free monocular 6-DoF estimation is feasible for long-range fixed-wing UAVs by combining heatmap-based anchoring, observation-ray conditioning, adaptive topological sampling, and geometry-constrained depth decoding. With a 4.96° mean rotation error, 82.8% Pose@10°/5%, and 4.6 ms inference—all without CAD models or external detections—the framework offers a favorable accuracy-prior-efficiency trade-off, and the accompanying FW-UAV6DPose dataset provides a standardized benchmark for this regime. The principal unresolved issue is validation beyond simulation.