Rock Instance Segmentation from Synthetic Images for Planetary Exploration Missions

As the complexity and operation distance of space missions rises, the demand of highly autonomous rovers increases as well. An aspect of autonomous rovers that has been specifically attracting much attention from the space community is semi-autonomous sampling from celestrial bodies. Detecting possible samples is important for their extraction, which is challenging due to the unstructured and unknown enviroment, and the lack of suitable datasets. This work addresses the task of sample collection in an unknown and unstructured environment by presenting a module for visual stone segmentation. Due to the limited training data for such scenarios, we apply a photo-realistic simulator to optimize an unknown instance segmentation network. We evaluate various manners of fine-tuning and show the positive effect of training on data highly related to the test data.