To support the development and evaluation of our drone pose estimation method, we constructed a high-quality synthetic dataset composed of two parts: 2DroneKey and 3DronePose. 

2DroneKey contains 2D drone images with annotated keypoints. It consists of 10 sequences (Seq.1 to Seq.10), each containing 1,000 frames, resulting in a total of 10,000 images. The drones used in this dataset are two popular models: DJI Air2S and DJI Mini2. All images are rendered at a resolution of 1920×1080 and feature realistic backgrounds captured with a 360-degree camera. While the translation of drones remains mostly fixed, a wide variety of rotations are included to help models learn robust keypoint localization under different poses. This dataset is primarily used for training and evaluating 2D keypoint detectors. 

3DronePose extends 2DroneKey by providing 6DoF pose annotations in addition to 2D keypoints. It includes three sequences (Seq.11 to Seq.13) with 500, 400, and 300 frames, respectively. Each sequence simulates different types of drone motion: Seq.11 features linear motion without rotation, Seq.12 includes nonlinear motion without rotation, and Seq.13 involves both nonlinear motion and rotation. The pose annotations were generated based on known drone geometry and virtual camera calibration. This dataset is used to evaluate the performance of 3D pose estimation methods, especially under challenging motion and viewing conditions.