| Prelegent: | Maciej Zięba, dr inż., Katedra Informatyki i Inżynierii Systemów, Wydział Informatyki i Zarządzania, Politechnika Wrocławska, e-mail: Maciej.Zięba@pwr.edu.pl. |
| Temat: | Deep generative models for 3D representations |
| Abstract: | During this talk, I am going to provide basic features of deep learning generative models for 3D representations. The main idea behind that generative model is to create the model capable to produce data from the true data distribution. In practical applications, where the data is represented by complex structures, like images, the problem of approximation true data distribution is a challenging task. The problem of training the generative models is well investigated in the literature. Usually, the deep generative models aim at transforming a random sample from the assumed prior distribution and transform the sample using a deep neural network to construct the sample from the data distribution. Various generative models are proposed in the literature like GANs, VAE or Normalizing Flows. The existing generative models are widely explored for images, text, and sound but there is still some there is field to explore for 3D representations like point clouds or meshes. During my talk, I am going to focus on novel deep generative models dedicated to 3D representations designed by me and my collaborators. I am going to focus on 3D Adverstrail Autoencoder that is a multitask generative model capable to generate 3D point clouds, but also it can cluster them and compress them to compact binary representations. Further, I am going to provide the details about flow-based generative models that are able to generate arbitrary given numbers of points and represent the generated object as a mesh. The presentation is going to be summarized with some quantitative and qualitative analysis of the designed algorithms. |
