DLVS3-HST Sample Dataset

Description

The DLVS3-HST dataset is a large-scale, physically-based synthetic image collection designed to advance deep learning and computer vision research in satellite pose estimation. It features over 640,000 high-resolution images of the Hubble Space Telescope (HST), each rendered with accurate lighting, realistic material properties, and comprehensive ground-truth annotations. The dataset includes precise 6-DoF pose information, 37 keypoint locations per image, and additional pixel-level data such as segmentation masks, depth maps, and surface normals. DLVS3 was developed to address the scarcity of real-world annotated space imagery and to support the development and benchmarking of robust, generalizable algorithms for autonomous spacecraft operations, rendezvous, servicing, and debris removal. The dataset leverages advanced rendering pipelines and domain randomization to ensure diversity and realism, hopefully making it a foundational resource for both academic research and industrial applications in vision-based navigation. A small sample of the dataset is available for direct download here. For access to the complete dataset (approx. 3.2 TB), please read the licence and register at: https://mi.services/dlvs3-hst-dataset-access/

For full technical details, methodology, and benchmarking results, please refer to our open-access white paper: Velkei, S. et al. "A large-scale, physically-based synthetic dataset for satellite pose estimation" arXiv:2506.12782 (2025). https://arxiv.org/abs/2506.12782.

For further information or collaboration inquiries, please contact: mind@mi.services

Number of images: 10000
Data source: Machine Intelligence DLVS3 Software

How to download

The sample dataset is freely available on Zenodo at the following link: https://zenodo.org/records/15731476.

References

• Velkei, S. et al. "A large-scale, physically-based synthetic dataset for satellite pose estimation" arXiv:2506.12782 (2025). https://arxiv.org/abs/2506.12782.