Abstract: The classical convolutional neural networks performance looks exceptionally great when the test dataset are very close to the training dataset. But when it is not possible, the accuracy of neural networks may even be reduced. The capsule networks are trying to solve the problems of the classical neural networks. Capsule networks are a brand new type of artificial neural networks, introduced by Geoffrey Hinton and his research team. In this work we would like to training capsule based neural networks for segmentation tasks, when the training set and test set are very different. For the training we use only computer generated virtual data, and we test our networks on real world data. We created three different capsule based architectures, based on classical neural network architectures, such as U-Net, PSP Net and ResNet. Experiences show how capsule networks are efficient in this special case.

Keywords: Capsule network, CapsNet, Neural network, Object segmentation, Virtual dataset.

.__________________________________________________________________________________________