ReSyRIS - A Real-Synthetic Rock Instance Segmentation Dataset for Training and Benchmarking

Abstract

The exploration of our solar system for understanding its creation and investigating potential chances of life on other celestial bodies is a fundamental drive of human mankind. After early telescope-based observation, Apollo 11 was the first space mission able to collect samples on the lunar surface and take them back to earth for analysis. Especially in recent years this trend accelerates again, and many successors were (or are in the process of being) launched into space for extra-terrestrial sample extraction. Yet, the abundance of potential failures makes these missions extremely challenging. For operations aimed at deeper parts of the solar system, the operational working distance extends even further, and communication delay and limited bandwidth increase complexity. Consequently, sample extraction missions are designed to be more autonomous in order to carry out large parts without human intervention. One specific sub-task particularly suitable for automation is the identification of relevant extraction candidates. While there exists several approaches for rock sample identification, there are often limiting factors in the form of applicable training data, lack of suitable annotations of the very same, and unclear performance of the algorithms in extra-terrestrial environments because of inadequate test data. To address these issues, we present ReSyRIS (Real-Synthetic Rock Instance Segmentation Dataset), which consists of real-world images together with their manually created synthetic counterpart. The real-world part is collected in a quasi-extra-terrestrial environment on Mt. Etna in Sicily, and focuses recordings of several rock sample sites. Every scene is re-created in OAISYS, a Blender-based data generation pipeline for unstructured outdoor environments, for which the required meshes and textures are extracted from the volcano site. This allows not only precise re-construction of the scenes in a synthetic environment, but also generation of highly realistic training data with automatic annotations in similar fashion to the real recordings. We finally investigate the generalization capability of a neural network trained on incrementally altered versions of synthetic data to explore potential sim-toreal gaps. The real-world dataset together with the OAISYS config files to create its synthetic counterpart are publicly available at https://rm.dlr.de/resyris_en. With this novel benchmark on extra-terrestrial stone instance segmentation we hope to further push the boundaries of autonomous rock sample extraction.