Random bin picking (RBP) has been a popular research topic due to the demand of industry 4.0, techniques like object detection, picking strategies and robot motion planning are more and more important. Much of the existing research uses the CAD model of the workpiece as the database. However, building CAD models is time-consuming and not all objects have CAD models. In this paper, a CAD-free random bin picking system is proposed to pick miscellaneous objects. By using Mask-RCNN instance segmentation, the object’s category and pickable area can be determined within a 2D image captured from RGB-D camera. Then, the pixels of pickable area can be converted into point clouds for picking tasks with the depth data of RGB-D camera. Compared with traditional RBP systems, a system with the Mask-RCNN doesn’t need to create CAD models, and it only requires fewer images of stacked objects (less than 50) and heuristic picking points labelling as the training data. Thus, the RBP systems which proposed in this paper can lowers the barriers to introduce the random bin picking system into factories. Through this scheme, a fast changeover for different objects could be made within 10 hours. The experiment results show that this system could pick two different objects with high success rate and acceptable cycle time. This system provides a useful and efficient solution for the industrial automation implementations that require bin picking.