Datasets:
tags:
- spacecraft
- pose_estimation
Multi-target Synthetic Dataset for Spacecraft Pose Estimation
Overview
This dataset was developed for 6D pose estimation of unseen, non-cooperative spacecraft in proximity-operations scenarios. Most existing datasets focus on a single target, which leads models to overfit to a specific spacecraft and limits their ability to generalize to previously unseen targets.
To address this limitation, the present dataset is multi-target and includes a wide variety of spacecraft geometries, sizes, and surface material properties. It also captures diverse viewing configurations and illumination conditions representative of realistic orbital environments. This diversity enables a more systematic and reliable evaluation of pose-estimation methods on previously unseen spacecraft.
The dataset was generated using a numerical tool designed for visible-spectrum imaging of spacecraft developed by the ASTRA Laboratory at Politecnico di Milano, and it is based on the methodology presented in this research.
Examples
Here are some examples of rendered images from the training dataset.
 |
 |
 |
 |
 |
 |
Dataset Format
The dataset is provided in the WebDataset format. Each sample consists of four files:
infos.json, which contains the scene and view identifiers for the sample. In this dataset, only the scene identifier is relevant, as each scene contains a single view.rgb.png, which stores the rendered RGB image.object_data.json, which includes the target pose expressed in the camera coordinate frame in the format [rotation quaternion, translation vector], along with the coordinates defining the 2D bounding box.camera_data.json, which specifies the image resolution and the camera intrinsic matrix K .
Dataset Composition
The dataset includes:
- 7,200 training samples
- 800 validation samples
- 2,000 test samples
Half of the images contain the Earth as background, while the remaining samples depict a deep-space background.
To prevent data leakage and ensure evaluation on unseen targets, the spacecraft models used for training and validation are disjoint from those used for testing.
Out of 28 spacecraft targets:
- 21 targets are used for training and validation
- 7 targets are reserved exclusively for testing
All spacecraft 3D meshes were sourced from the ESA public repository. The selected test spacecraft cover a broad range of appearance characteristics, including:
- symmetric vs. asymmetric structures
- varying prominence of appendages (e.g., solar panels and antennas) relative to the main bus
- different surface reflectance properties and contrast levels against a dark background
This heterogeneity supports a meaningful and generalizable evaluation of pose-estimation algorithms.
Here's a complete list of the targets employed.
| Training+Validation | Test |
|---|---|
| BepiColombo | ISO |
| Cassini–Huygens | Proba-3 OSC |
| CHEOPS | Solar Orbiter |
| Cluster | Trace Gas Orbiter |
| Double Star | Venus Express |
| Envisat | XMM-Newton |
| Euclid | Ulysses |
| Gaia | |
| Giotto | |
| Herschel | |
| Hubble Space Telescope | |
| INTEGRAL | |
| JUICE | |
| LISA Pathfinder | |
| Mars Express | |
| Planck | |
| Proba-2 | |
| Proba-3 CSC | |
| Rosetta | |
| SMART-1 | |
| SOHO |