6DroneSyn: Synthetic Benchmark for Drone Pose Estimation

• 6DroneSyn is a synthetic benchmark for monocular drone 3D pose estimation that addresses prior limitations in scale, model diversity, and annotation completeness.
• It integrates real 360° panoramic imagery with synthetic drone rendering to deliver high-resolution, sequence-based data and dense 2D/3D annotations.
• The benchmark enables rigorous evaluation of pose estimation methods by measuring rotation and translation errors across multiple DJI drone models.

6DroneSyn is a synthetic benchmark for monocular drone 3D pose estimation introduced in connection with the DroneKey++ framework. It is defined as a large-scale, high-resolution, sequence-based dataset intended to address limitations of earlier drone pose corpora in scale, model diversity, environmental realism, and annotation completeness. The benchmark contains 52,920 frames at $1920\times1080$, spans 7 DJI drone models and 88 outdoor backgrounds, and provides per-frame 2D and 3D supervision including bounding boxes, keypoints, rotation, translation, camera intrinsics, class labels, and full sequence trajectories (Hwang et al., 5 Feb 2026). Its design is explicitly motivated by the difficulty of obtaining precise 3D pose labels in real outdoor scenes and by the need for a benchmark that better supports generalization across drone types and surveillance-like environments (Hwang et al., 5 Feb 2026).

1. Definition and research motivation

6DroneSyn was created to overcome three limitations identified in prior drone pose estimation datasets: limited scale and diversity, restricted environmental realism and variety, and the practical difficulty of collecting accurate 3D pose labels in real outdoor scenes (Hwang et al., 5 Feb 2026). Earlier datasets cited in the underlying study include single-model or low-diversity corpora such as DronePose, UAV-ADT, and 3DKey+3DPose. In this context, 6DroneSyn is positioned as a purpose-built benchmark for single-camera drone 3D pose estimation, while also being intended as a general benchmark for subsequent work (Hwang et al., 5 Feb 2026).

The dataset is synthetic throughout, but its visual generation process is grounded in real 360-degree environment captures rather than purely virtual backgrounds. This design choice aims to reduce the domain gap relative to CCTV-like or surveillance imagery by embedding rendered drone models into panorama-derived outdoor scenes (Hwang et al., 5 Feb 2026). The source paper further states that all publicly available drone pose estimation datasets are synthetic, because obtaining accurate full 3D pose annotations under real outdoor conditions is “nearly impossible” due to regulations, equipment, and cost (Hwang et al., 5 Feb 2026).

Within the DroneKey++ study, 6DroneSyn serves a dual role. It is both the only dataset used for training and evaluation of the proposed prior-free pose estimation framework and the benchmark on which comparisons are performed against Keypoint Detector (YOLOv8) + PnP, DroneKey, and DronePose re-implemented without size or mesh priors (Hwang et al., 5 Feb 2026). This tightly coupled usage suggests that 6DroneSyn was engineered not merely as a repository of images, but as a benchmark whose annotation schema aligns directly with a multi-task monocular pose estimation pipeline.

2. Scale, composition, and scene organization

6DroneSyn contains 52,920 images at a resolution of $1920\times1080$, with an on-disk size of about 120 GB (Hwang et al., 5 Feb 2026). The data are organized into 13 scenes and 91 sequences. Each scene contains 21 subsequences, where each subsequence corresponds to one drone type and one background. The scene structure yields $13 \times 21 = 91$ sequences in total (Hwang et al., 5 Feb 2026).

The temporal composition of the dataset is specified at the scene level. Scenes #01–#03 contain linear motion, 21 subsequences, 30 FPS, and 4 seconds per subsequence, producing 120 frames per subsequence and 7,560 frames in total across these scenes. Scenes #04–#06 also use 30 FPS and 4-second subsequences, but with non-linear motion, contributing another 7,560 frames. Scenes #07–#09 return to linear motion with the same total of 7,560 frames. Scenes #10–#13 are non-linear, 30 FPS, and 12 seconds per subsequence, producing 360 frames per subsequence and 30,240 frames overall. Summed together, these groups produce the full 52,920-frame corpus (Hwang et al., 5 Feb 2026).

The benchmark includes 7 DJI drone models: Mini3 Pro, Mini2, Air3, Air2, Mavic 2 Pro (Mav2), Mavic3 (Mav3), and Tello (Hwang et al., 5 Feb 2026). This model diversity is an explicit part of the benchmark design. The source study states that DroneKey++ relies on class embeddings to encode physical size and shape implicitly, so the dataset must include models with different geometries and scales for the class embedding to be meaningful (Hwang et al., 5 Feb 2026). A plausible implication is that 6DroneSyn is not only a benchmark for pose recovery, but also a benchmark for cross-model geometric generalization.

The environmental side of the dataset consists of 88 outdoor backgrounds derived from 22 real 360-degree environments. Each environment is rotated around the $Z$-axis by 90-degree increments, producing four background variations per environment and thus $22\times 4=88$ backgrounds (Hwang et al., 5 Feb 2026). In each scene, 7 drone types are paired with 3 background variations, resulting in the 21 subsequences per scene (Hwang et al., 5 Feb 2026).

3. Synthetic generation pipeline and trajectory design

The generation pipeline for 6DroneSyn combines real 360-degree panoramic imagery with synthetic drone rendering (Hwang et al., 5 Feb 2026). First, 22 real outdoor environments are captured as 360-degree images. Each panorama is then rotated by $90^\circ$ increments around the vertical axis, yielding views at $\theta=0^\circ,90^\circ,180^\circ,270^\circ$ (Hwang et al., 5 Feb 2026). This rotation augmentation increases appearance diversity while decorrelating scene layout from drone pose.

For each frame, a perspective camera view is rendered from the panorama. The paper does not specify the rendering engine, but it does specify the conceptual structure: a camera orientation and field of view are sampled, and the 360-degree panorama is projected into a standard perspective image of $1920\times1080$ (Hwang et al., 5 Feb 2026). A synthetic 3D drone model is then rendered as the foreground object at a defined 3D pose relative to the camera.

The dataset is explicitly sequence-based rather than an unordered collection of still images. Scenes #01–#03 and #07–#09 use linear motion trajectories, while scenes #04–#06 and #10–#13 use non-linear motion trajectories, including complex 3D paths (Hwang et al., 5 Feb 2026). The source text states that these trajectories range “from simple linear movements to complex 3D trajectories, encompassing all types of drone motions in real operations” (Hwang et al., 5 Feb 2026). This wording is broad, but it clearly establishes that temporal pose evolution is a first-class component of the benchmark.

The benchmark defines 3D pose as drone pose in the camera coordinate system. The transformation is written as

$$\begin{bmatrix} \mathbf{x}_c \ 1 \end{bmatrix} = \begin{bmatrix} R & \mathbf{t} \ 0 & 1 \end{bmatrix} \begin{bmatrix} \mathbf{x}_d \ 1 \end{bmatrix},$$

where $\mathbf{x}_d$ is a point in the drone’s local coordinate frame, $1920\times1080$0 and $1920\times1080$1 are the drone-to-camera rotation and translation, and $1920\times1080$2 is the corresponding point in camera coordinates (Hwang et al., 5 Feb 2026). In 6DroneSyn, $1920\times1080$3 and $1920\times1080$4 are part of the ground-truth annotation set.

This trajectory-centered design is significant because it enables evaluation not only of per-frame monocular inference but also of sequential smoothing, temporal consistency, and full 3D trajectory reconstruction. The DroneKey++ study notes that sequences are used for temporal visualization, trajectory analysis, and Gaussian smoothing of predicted poses (Hwang et al., 5 Feb 2026). This suggests that 6DroneSyn occupies an intermediate position between a classical image dataset and a video benchmark.

4. Annotation schema and geometric supervision

A defining characteristic of 6DroneSyn is its dense annotation set. Each frame includes 2D bounding boxes, 2D keypoints, 3D keypoints, 3D pose, camera intrinsics, a drone class label, and sequence-level temporal ordering (Hwang et al., 5 Feb 2026). The annotations are automatically generated within the rendering pipeline, which the source paper states “completely eliminat[es] human labeling errors” (Hwang et al., 5 Feb 2026).

The 2D annotations consist first of bounding boxes $1920\times1080$5 for the drone region in the full frame (Hwang et al., 5 Feb 2026). The second 2D annotation type is a set of four keypoints corresponding to the propellers of the quadcopter. These are represented in the paper as

$1920\times1080$6

and the predicted 2D keypoints are defined by

$1920\times1080$7

The dataset provides the ground truth for $1920\times1080$8 (Hwang et al., 5 Feb 2026).

The 3D keypoint annotation comprises the corresponding 3D coordinates of the same four propellers in camera coordinates:

$1920\times1080$9

The 3D pose decoder in DroneKey++ is trained to regress these points using an MLP and an MSE loss against the provided $13 \times 21 = 91$0 (Hwang et al., 5 Feb 2026). This means that 6DroneSyn supports both explicit pose supervision through $13 \times 21 = 91$1 and $13 \times 21 = 91$2 and structural supervision through sparse 3D part geometry.

The pose annotation itself consists of rotation and translation. Rotation is stored as angles around axes, normalized to $13 \times 21 = 91$3; the network predicts a 3-vector $13 \times 21 = 91$4 after sigmoid, with ground truth $13 \times 21 = 91$5 normalized in the same way (Hwang et al., 5 Feb 2026). Translation is a 3D vector representing the drone center in camera coordinates,

$13 \times 21 = 91$6

and translation supervision uses MSE loss:

$13 \times 21 = 91$7

Camera intrinsics are supplied as $13 \times 21 = 91$8, enabling back-projection from 2D keypoints to rays:

$13 \times 21 = 91$9

The class label is one of the seven DJI models and is used to supervise a 7-way classification head (Hwang et al., 5 Feb 2026). Finally, full sequence trajectory annotations include time-ordered 2D keypoints, 3D keypoints, rotation, and translation for all frames in a sequence (Hwang et al., 5 Feb 2026).

The richness of this annotation schema is central to the benchmark’s research value. It permits monocular pose estimation to be formulated as a joint problem over appearance, part structure, geometric rays, class-conditioned size and shape, and temporal continuity. This differs from datasets that supply only boxes, only 2D keypoints, or only sparse pose metadata.

5. Evaluation protocol and role in DroneKey++

The official split in 6DroneSyn is scenario-wise rather than frame-wise. Training uses Scenes 01–02, 04–05, 08, and 10–12; validation uses Scenes 03, 09, and 13; testing uses Scenes 06–07 (Hwang et al., 5 Feb 2026). Because test scenes are never seen during training or validation, the protocol evaluates generalization across held-out motion scenarios and background combinations rather than random temporal subsampling. The source text does not indicate that any drone model is held out completely; instead, all 7 models appear across train, validation, and test at the dataset level (Hwang et al., 5 Feb 2026).

Evaluation is conducted in terms of rotation and translation errors. Rotation error is derived from the relative rotation between predicted and ground-truth rotation matrices:

$Z$0

followed by the angular distance

$Z$1

The benchmark reports both mean absolute error and median absolute error for rotation:

$Z$2

$Z$3

In practice, the paper reports rotation error in degrees (Hwang et al., 5 Feb 2026).

Translation error is defined as Euclidean distance between predicted and ground-truth translations:

$Z$4

with

$Z$5

These are reported per drone type and per scene in the source paper’s Table 3 and then averaged (Hwang et al., 5 Feb 2026).

6DroneSyn is structurally aligned with DroneKey++, whose supervision is decomposed into encoder and decoder losses:

$Z$6

$Z$7

$Z$8

The dataset provides exactly the labels required by this training decomposition: 2D keypoints, drone class, 3D keypoints, rotation, translation, and camera intrinsics (Hwang et al., 5 Feb 2026). The benchmark therefore functions simultaneously as an evaluation standard and as a dataset whose annotation semantics encode the architectural assumptions of the associated method.

6. Empirical findings, comparative position, and limitations

On the 6DroneSyn test scenes #06 and #07, DroneKey++ achieves a rotation MAE of $Z$9, rotation MedAE of $22\times 4=88$0, translation MAE of $22\times 4=88$1, and translation MedAE of $22\times 4=88$2 (Hwang et al., 5 Feb 2026). These are averages over all 7 drone models and both test scenes. Inference speed is reported as 19.25 FPS on CPU and 414.07 FPS on GPU (Hwang et al., 5 Feb 2026). Within the same benchmark, the paper compares against DronePose re-implemented without mesh priors, DroneKey + PnP, and a generic keypoint detector (YOLOv8) + PnP (Hwang et al., 5 Feb 2026).

The authors describe 6DroneSyn as challenging. Even with synthetic images and exact annotations, the mean rotation error remains on the order of $22\times 4=88$3, which they interpret as evidence of realistic pose variation and perspective difficulty (Hwang et al., 5 Feb 2026). Ablation studies performed entirely on 6DroneSyn further show that removing the keypoint encoder and instead using ground-truth 2D keypoints and class labels yields substantially worse rotation MAE, $22\times 4=88$4 instead of $22\times 4=88$5 (Hwang et al., 5 Feb 2026). The paper states that this suggests end-to-end learning from images, with 2D and 3D supervision, is important. It also reports that the class embedding enabled by the 7 drone classes reduces translation MAE from $22\times 4=88$6 to $22\times 4=88$7 (Hwang et al., 5 Feb 2026).

Relative to earlier datasets, 6DroneSyn occupies a distinctive position in the design space.

Dataset aspect	6DroneSyn	Comparative context
Frames	52,920	Similar scale to DronePose’s 55,988
Classes	7 models	More than DronePose and 3DKey+3DPose
Backgrounds	88 outdoor backgrounds	Derived from real 360° environments
Resolution	$22\times 4=88$8	High-resolution synthetic data
Annotations	2D(B), 2D(K), 3D(R+t)	Matches the richest prior pose annotation level

The source paper also reports feature-distribution analysis using PCA and t-SNE, comparing real web images, existing synthetic datasets, and 6DroneSyn. Existing synthetic data appear tightly clustered, whereas real images and 6DroneSyn are more broadly distributed; in t-SNE, 6DroneSyn overlaps real clusters, which the authors interpret as suggesting reduced domain gap (Hwang et al., 5 Feb 2026). This is an empirical claim about representation-space similarity rather than a direct deployment evaluation, but it reinforces the benchmark’s intended role as a visually more realistic synthetic corpus.

Several limitations are explicitly acknowledged. 6DroneSyn is synthetic only, even though the panoramas are real; no real drone images are included (Hwang et al., 5 Feb 2026). The benchmark currently covers only 7 DJI quadcopters, excluding other manufacturers and form factors such as fixed-wing drones and VTOL hybrids (Hwang et al., 5 Feb 2026). Environmental diversity, while substantial, does not fully include extreme weather, night, or indoor-outdoor transitions (Hwang et al., 5 Feb 2026). The current annotation design also assumes four visible propellers and does not include explicit occlusion labels or visibility flags, motivating future work on keypoint confidence mechanisms and improved robustness under occlusion (Hwang et al., 5 Feb 2026). Finally, the benchmark is monocular and RGB-only, with no multi-view, depth, infrared, radar, or LiDAR modalities (Hwang et al., 5 Feb 2026).

These limitations are consequential for interpretation. 6DroneSyn should not be understood as a complete benchmark for all UAV perception problems; rather, it is a benchmark optimized for monocular RGB 3D pose estimation in surveillance-like outdoor imagery. Its main contribution lies in unifying high resolution, multiple drone classes, panorama-derived outdoor diversity, complete 2D and 3D annotations, and temporal sequence structure within a single public synthetic benchmark (Hwang et al., 5 Feb 2026). A plausible implication is that future research may use it both as a standalone testbed for monocular pose inference and as a pretraining or controlled-ablation resource for broader drone perception systems.