Research projects

During rover missions to Mars the identification of promising scientific targets is a task with a high effort carried out by experienced planetary scientists. They determine locations a rover will visit to take a closer look, perform diverse measurements and eventually take probes. It would be a great advantage if rovers can identify scientific targets autonomously, supporting scientists with strategic decision-making by preselecting possibly interesting regions. Thereby, it should increase the overall scientific discoveries and help to avoid overlooking interesting features on the Martian surface.     

The project Mars-DL led by JOANNEUM Research investigates in collaboration with SLR Engineering and the Museum of Natural History of Vienna if this goal can be achieved by a deep learning (DL) system. DL systems proofed to be highly successful for object and pattern recognition. But they need huge amounts of training data to work reliably. Since this amount of data is not available from past and ongoing rover missions it needs to be simulated.

This is the task of VRVis in this project. We extended PRo3D, our viewer for the exploration and geologic interpretation of planetary surface reconstructions, to automatically render different types of training images. It was decided to train the system to detect Shatter Cones, which are created by asteroid impacts. Different SC models are positioned automatically on a chosen Martian surface area. Their sizes and orientations are varied. Furthermore, they are colour adjusted and shaded from the same illumination directing as the surface to perfectly blend into the scene.  PRo3D is controlled with a script to automatically render images from different viewpoints of the simulated scene. It can also generate masks and depth images, which are also necessary for training.

Currently the project is in the phase of image generation and training.