The full repository will come here soon!
Through the use of novel deep learning architectures the computer vision community achieved remarkable results on classification of 3D objects represented in point cloud data format. However, there has been hardly any work using this in the Zero-shot learning setting. In this paper we explore a novel approach to handle the 3D object classification task in the zero-shot setting. Instead of taking only one representative for each class we use several textual descriptions for each point cloud object and embed these into the same space. We demonstrate that, on our dataset with three unseen classes, this bypasses the problem of only covering a small part of the embedding space with the class representatives and therefore shows better and more stable results than the current state-of-the-art approach.