ConsInstancy: learning instance representations for semi-supervised panoptic segmentation of concrete aggregate particles

Abstract We present a semi-supervised method for panoptic segmentation based on ConsInstancy regularisation, novel strategy learning. It leverages completely unlabelled data by enforcing consistency between predicted instance representations and semantic segmentations during training in order to improve the performance. To this end, we also propose new types of that can be one simple forward path through fully convolutional network (FCN), delivering convenient simple-to-train framework segmentation. More specifically, prediction three-dimensional orientation map as intermediate representation two complementary distance transform maps final representation, providing unique test our challenging sets both, hardened fresh concrete, latter being proposed authors paper demonstrating effectiveness approach, outperforming results achieved state-of-the-art methods In particular, are able show leveraging approach overall accuracy (OA) is increased up 5% compared an entirely supervised using only labelled data. Furthermore, exceed OA 1.5%.