Aura-VisionPose: Vision-Based Satellite Pose Dataset

This dataset contains over 60000 labeled images of NASA's Aura satellite. Each image is a PNG showing a computer-generated rendering of the satellite orbiting over Earth, with a corresponding JSON file providing its pose with relation to the viewer. This was generated using Blender to allow us to create a series of images with different poses, realistic lighting, and backgrounds consistent with the rest of the image. The intended purpose of these images is to be used as training data for a visual navigation model. A second folder contains 1000 images of the same satellite, but with random images as backgrounds to prevent overfitting. 

Each image is labeled with the relative position (meters), relative orientation (quaternions), and bounding box coordinates. The camera frame follows a right-handed convention: +x points to the right in the image, +y points upward, and +z points behind the camera. Bounding boxes are provided in pixel coordinates for 1920 × 1080 images, with the origin at the bottom-left corner. The satellite position is referenced to the center of the satellite bus. The target body frame is defined such that +x points opposite the solar panel direction, +y points toward the satellite dish, and +z completes the right-handed coordinate system. The images are taken with a camera with an 80° horizontal FOV.

The dataset is part of DCV-Space group's ongoing efforts to develop a multi-agent rendezvous and proximity operations framework (which includes a testbed). To cite the paper corresponding to this dataset, please use the following MMEDR-Autonomous framework paper: https://doi.org/10.48550/arXiv.2603.20575