The paper discusses the results of the investigation of the rock chunks vein segmentation task using deep-learning computer vision systems. The previously collected and labeled database of asbestos veins in the rock chunks of the Open Pit is under study. The task has practical novelty, and allows carrying out real-time-scale automatic productivity estimation of the open pit. From the reseapoint of of view field, to the best of our knowledge, the task is mostly related to the fragmentation one. However, it is shown that only base-line solutions (using deep learning) here have been researched. Most of those solutions using either a semantic segmentation approach with further instances division or strictly instance segmentation approach. The current paper investigate the last one comparing several Mask-R-CNN architectures. The usage of the methods under research in the described area draw up the novelty of the study. The current results provides the base-line solution for the next researches in this field.