The speed of adopting new technologies in industrial automation depends on two factors, reliability and ease of integration. CNN-based object segmentation is one of those technologies that are well developed in research and other industries but still not well established in industrial automation. It is an essential processing step for robotic grasping. Nevertheless, most of the grasping in the industry is still computed by classical non-learning algorithms or based on simple manually programmed hypotheses. The traditional setup in most research related to the object segmentation problem is to have a finite number of objects/classes. While this is suitable for some other problems, it is the hurdle stopping the ease of integrating CNN object segmentation in the industry. A more practical approach is to use object class-agnostic segmentation, where a CNN is used to segment objects in an image without classifying them. Then classical feature extractors can be used for the classification process. This method would avoid the need for manual tailoring of CNNs for each individual setup/environment. In this work, we propose an image processing pipeline that is general and invariant to setup. We also show the feasibility of the class-agnostic segmentation, discuss the feasibility of using purely synthetic data for the CNN training and its results when deployed and tested on our real setup.
Object class-agnostic segmentation for practical CNN utilization in industry
Object class-agnostic segmentation for practical CNN utilization in industry.