DroneKey++: Prior-Free Drone Pose Estimation

• The paper introduces a novel prior-free framework that jointly predicts 2D keypoints, class labels, and 3D drone pose from monocular imagery.
• It leverages a transformer-based encoder and ray-based geometric reasoning to mitigate monocular scale ambiguity without explicit size priors.
• The approach is validated on the synthetic 6DroneSyn benchmark, achieving significant improvements in real-time rotation and translation accuracy.

Searching arXiv for the primary paper and closely related drone pose-estimation work to ground the article. DroneKey++ is an end-to-end, size-prior-free framework for estimating the full $6$-DoF $3$D pose of drones from single monocular images, evaluated on sequential data, together with a synthetic benchmark named 6DroneSyn (Hwang et al., 5 Feb 2026). It removes the need for manually supplied information about a drone’s physical size or CAD mesh and instead jointly performs $2$D keypoint detection, drone classification, and $3$D pose estimation within a single network. The method is organized around a keypoint encoder and a pose decoder; the former predicts propeller keypoints and class probabilities, while the latter uses ray-based geometric reasoning and class embeddings to infer rotation and translation. The accompanying dataset is designed to address the limited scale, model diversity, and environmental variability of earlier drone pose resources (Hwang et al., 5 Feb 2026).

1. Problem formulation and motivation

DroneKey++ addresses monocular drone pose estimation in the camera coordinate frame. Given an RGB image, or a sequence processed frame by frame, the goal is to estimate the drone’s rotation $\mathbf{R} \in SO(3)$ and translation $\mathbf{t} \in \mathbb{R}^3$. In the formulation used by the paper, this pose determines the drone’s line of sight, potential target direction, and $3$D trajectory over time when estimates are concatenated across frames (Hwang et al., 5 Feb 2026).

The framework is motivated by anti-drone and surveillance settings in which accurate $3$D pose is operationally more informative than $2$D detection alone. Previous methods are described as depending on physical size priors, such as known distances between propellers for PnP, or on $3$D mesh priors used during rendering-based training, while existing datasets are characterized as small-scale, single-model, or constrained in environment diversity (Hwang et al., 5 Feb 2026). This combination makes generalization difficult and limits deployment in scenarios where an operator does not know the exact model, physical dimensions, or mesh of an unauthorized drone.

The central difficulty is monocular scale ambiguity: the same $3$0D projection can correspond to a small nearby drone or a larger distant drone. DroneKey++ defines “size prior-free” as the absence, at inference time, of explicit physical scale or $3$1D model information. Instead, the model infers orientation and absolute translation from images, camera intrinsics, and learned parameters. A plausible implication is that the framework replaces explicit metric structure with a learned coupling between appearance, class-dependent morphology, and geometric rays (Hwang et al., 5 Feb 2026).

2. Overall framework and representational structure

The architecture is divided into two modules. The keypoint encoder takes an RGB image and outputs $3$2D keypoints $3$3, representing the four propeller centers, and a drone class distribution $3$4, where $3$5 on 6DroneSyn. The encoder also produces a global $3$6 token that serves as a class embedding for the decoder (Hwang et al., 5 Feb 2026).

The pose decoder consumes predicted $3$7D keypoints, class probabilities, and the camera intrinsic matrix $3$8. It computes $3$9D rays, embeds both rays and class information, predicts relative $2$0D keypoints $2$1, and regresses a $2$2-dimensional pose vector $2$3, which is decomposed into rotation and translation (Hwang et al., 5 Feb 2026).

Training is joint. The paper defines a combined objective

$2$4

with the encoder supervising $2$5D keypoints and class labels, and the decoder supervising $2$6D keypoints, translation, and rotation (Hwang et al., 5 Feb 2026). The paper reports that equal weighting of these losses performed better than alternative weighting schemes.

This organization is significant because it makes class prediction part of the pose-estimation pipeline rather than an auxiliary side task. The reported interpretation is that class probabilities implicitly encode scale and shape cues, which are then used by the decoder to mitigate monocular ambiguity without external priors (Hwang et al., 5 Feb 2026).

3. Keypoint encoder and multi-task supervision

The keypoint encoder extends the transformer-based encoder of DroneKey with a ViT-style $2$7 token for classification. A ResNet backbone first extracts feature maps

$2$8

which are partitioned into patches and flattened into token embeddings

$2$9

with positional encodings added (Hwang et al., 5 Feb 2026). The $3$0 token is prepended, and the transformer encoder applies $3$1 self-attention layers.

For keypoint regression, DroneKey++ uses gated multi-layer aggregation. At each transformer layer $3$2, the feature tokens are max-pooled across spatial positions to obtain an intermediate representation $3$3, then linearly projected to $3$4 image-plane keypoints: $3$5 A gate vector is predicted from the last intermediate representation,

$3$6

and the final keypoints are obtained by gated summation followed by ReLU: $3$7 ReLU is used to avoid negative coordinates (Hwang et al., 5 Feb 2026).

Classification uses the final $3$8 token: $3$9 The encoder loss is the sum of MSE for $\mathbf{R} \in SO(3)$0D keypoints and cross-entropy for class prediction: $\mathbf{R} \in SO(3)$1 No weighting factors are applied; the paper states that equal weights worked best (Hwang et al., 5 Feb 2026).

The encoder is therefore not only a detector of propeller centers but also a learned scene-level summarizer. This helps explain an ablation result in which the learned encoder outperformed a setting that fed ground-truth keypoints and class labels directly to the decoder. The paper interprets this as evidence that the encoder learns contextual features that pre-condition downstream pose regression (Hwang et al., 5 Feb 2026).

4. Pose decoder, ray-based geometry, and the sequential-image setting

The pose decoder uses explicit projective geometry. For each keypoint $\mathbf{R} \in SO(3)$2 and camera intrinsics $\mathbf{R} \in SO(3)$3, the decoder computes an unnormalized ray direction

$\mathbf{R} \in SO(3)$4

Each ray is embedded into $\mathbf{R} \in SO(3)$5 by a linear layer, producing $\mathbf{R} \in SO(3)$6, while class probabilities are projected into a class embedding $\mathbf{R} \in SO(3)$7. These are concatenated,

$\mathbf{R} \in SO(3)$8

and used by two MLP heads (Hwang et al., 5 Feb 2026).

The first head estimates relative $\mathbf{R} \in SO(3)$9D propeller coordinates in a drone-centric frame: $\mathbf{t} \in \mathbb{R}^3$0 The second uses both fused features and predicted $\mathbf{t} \in \mathbb{R}^3$1D keypoints: $\mathbf{t} \in \mathbb{R}^3$2 The first three output dimensions, after sigmoid, form a normalized rotation representation

$\mathbf{t} \in \mathbb{R}^3$3

and the final three are the translation

$\mathbf{t} \in \mathbb{R}^3$4

The paper emphasizes that this normalized angle representation is used for regression only; evaluation uses full rotation matrices (Hwang et al., 5 Feb 2026).

Decoder supervision includes MSE on $\mathbf{t} \in \mathbb{R}^3$5D keypoints, MSE on translation, and a circular loss for normalized angles that uses the shortest distance on the unit circle. The decoder loss is

$\mathbf{t} \in \mathbb{R}^3$6

Ablation results reported in the paper show that ray embedding strongly improves translation error and that adding class embeddings produces the best overall decoder, particularly for translation MAE (Hwang et al., 5 Feb 2026). This suggests that ray geometry contributes projective structure, while class embeddings provide implicit scale and morphology cues.

Despite the title’s reference to sequential images, the method itself is frame-wise. The network contains no explicit temporal model, no recurrent state, and no temporal fusion. Sequences are used in dataset organization, evaluation, and visualization, and trajectories are formed by concatenating per-frame estimates, sometimes with Gaussian smoothing (Hwang et al., 5 Feb 2026). A common misconception is therefore that DroneKey++ is a temporal pose-estimation network; in the reported implementation it is not.

5. 6DroneSyn benchmark dataset

DroneKey++ is introduced together with 6DroneSyn, a synthetic benchmark for drone $\mathbf{t} \in \mathbb{R}^3$7D pose estimation. The dataset contains $\mathbf{t} \in \mathbb{R}^3$8 high-resolution images, approximately $\mathbf{t} \in \mathbb{R}^3$9 GB in total, at $3$0 resolution. It covers $3$1 DJI drone models—Mini3 Pro, Mini2, Air3, Air2, Mavic 2 Pro, Mavic3, and Tello—and $3$2 outdoor backgrounds generated from $3$3 real $3$4 panoramas, each rotated $3$5 around the $3$6-axis to yield four variants (Hwang et al., 5 Feb 2026).

The benchmark comprises $3$7 synthetic scenes. Each scene contains $3$8 subsequences corresponding to $3$9 drone types and $3$0 backgrounds. Scenes #01–#03 and #07–#09 are linear, scenes #04–#06 and #10–#13 are non-linear, and the total frame counts are $3$1, $3$2, $3$3, and $3$4, respectively, with durations of $3$5 s for the first three groups and $3$6 s for the last group at $3$7 FPS (Hwang et al., 5 Feb 2026). The motion patterns include linear and complex $3$8D trajectories such as ascending, descending, and turning.

Annotations include $3$9D bounding boxes, $2$0D keypoints for the four propellers, $2$1D keypoints, full $2$2D pose as rotation vectors, rotation matrices, and translation, as well as class labels and camera intrinsics and extrinsics (Hwang et al., 5 Feb 2026). Because labels are renderer-generated, the paper characterizes them as noise-free relative to manual annotation.

The benchmark is explicitly framed as a response to the limitations of earlier drone datasets. The paper contrasts 6DroneSyn with UAVA (DronePose), UAV-ADT, and 3DKey+3DPose in terms of model diversity, scale, and environmental variation (Hwang et al., 5 Feb 2026). It also reports PCA and t-SNE analyses in which features extracted from 6DroneSyn overlap more broadly with real web-collected drone frames than those from an existing synthetic dataset, suggesting reduced domain gap and greater visual diversity. This suggests that 6DroneSyn is intended not only as a training corpus but also as a stress test for cross-model generalization.

6. Empirical performance, ablations, and limitations

All reported experiments are conducted on 6DroneSyn with a scenario-wise split: training on scenes 01–02, 04–05, 08, and 10–12; validation on scenes 03, 09, and 13; and testing on scenes 06 and 07 (Hwang et al., 5 Feb 2026). This split is designed to avoid frame-level leakage and to test generalization across backgrounds and motion patterns. Evaluation uses MAE and MedAE for both rotation and translation, with rotation error computed from the minimal angle between predicted and ground-truth rotation matrices and translation error computed as Euclidean distance between translations.

On the test set, DroneKey++ reports average rotation MAE $2$3 and MedAE $2$4, and translation MAE $2$5 and MedAE $2$6 (Hwang et al., 5 Feb 2026). The paper compares this with a prior-free reimplementation of DronePose, DroneKey, and a keypoint detector plus PnP baseline, stating that DroneKey++ improves rotation MAE by about $2$7 over DronePose and about $2$8 over DroneKey, and improves translation MAE by about $2$9 over DronePose and about $3$0 over DroneKey. Qualitative results are described as producing smooth and accurate trajectories and orientations across different drones and scenes.

The framework is also explicitly evaluated for throughput. Reported inference speeds are $3$1 FPS on an NVIDIA A100 GPU and $3$2 FPS on an Intel Xeon Gold 6440 CPU, including both encoder and decoder in a single forward pass and without external PnP or optimization (Hwang et al., 5 Feb 2026). The paper therefore characterizes the method as suitable for real-time surveillance scenarios.

Ablation studies isolate several design choices. Replacing the learned encoder with direct ground-truth keypoints and class labels worsens performance substantially, with rotation MAE changing from $3$3 to $3$4 and translation MAE from $3$5 to $3$6 (Hwang et al., 5 Feb 2026). Decoder ablations show that ray embedding greatly improves translation relative to a direct MLP pose head, and that the full decoder with both ray and class embeddings yields the best overall result, including translation MAE $3$7. Equal loss weighting also outperforms tanh-weighted, smoothly shifted, and $3$8D-biased alternatives.

Training uses an ImageNet-pretrained ResNet backbone, Adam optimizer with initial learning rate $3$9 and cosine annealing, batch size $3$00, and $3$01 epochs; no data augmentation is applied in the reported experiments (Hwang et al., 5 Feb 2026). This is noteworthy because it places the reported generalization on synthetic domain design and multi-task supervision rather than augmentation-based regularization.

The paper identifies several limitations. Real-world validation remains limited despite the use of panorama-based synthesis to reduce synthetic-to-real discrepancy. The decoder assumes reliable keypoints and therefore may degrade under occlusion or motion blur. The benchmark is restricted to $3$02 DJI models and $3$03 backgrounds, and the current formulation assumes a single drone per image. Temporal modeling is absent, with trajectory stabilization handled only by post-processing (Hwang et al., 5 Feb 2026). Future directions explicitly proposed include confidence-aware keypoint prediction, richer real-world evaluation, more drone and environment types, multi-drone scenes, and explicit temporal models such as transformers over time or Kalman filters.

Within the literature on drone pose estimation, DroneKey++ is defined by the conjunction of three elements: prior-free monocular $3$04-DoF estimation, joint keypoint-class-pose learning, and a synthetic benchmark designed to test generalization across drone models and outdoor contexts (Hwang et al., 5 Feb 2026). Its principal methodological claim is that class embeddings combined with ray-based geometric reasoning can substitute for explicit size or mesh priors in metric pose regression.