Capsule-based Autoencoder Network for Pointcloud Reconstruction

The theory of capsule networks and the dynamic routing mechanism for capsules was introduced by Geoffrey Hinton and his research team. In this new approach, they tried to solve typical problems of classical convolutional neural networks. For example, that the efficiency of neural networks degrades when a geometric transformation is applied on the input image, or when the data is far away from the training dataset. It became clear early on that capsule networks are state-of-the-art solutions for visual data classification tasks. For other tasks their use is less common and in many cases difficult to apply. For example image segmentation or object detection and localization. The efficiency of the capsule networks theory in the field of pointcloud processing is also an open question. In this work we investigated the pointcloud reconstruction capability of capsule networks. In this approach, three different complexity autoencoder networks was selected. We created a decoder network based on capsules theory, which was fitted to the existing autoencoder networks. The efficiency of the networks was tested using four different datasets. As a result of our work, we show the effectiveness of capsule networks in the field of pointcloud reconstruction compared with the selected autoencoder networks.