Learning a Hierarchical Latent-Variable Model of Voxelized 3D Shapes

We propose the Variational Shape Learner (VSL), a hierarchical latent-variable model for 3D shape learning. VSL employs an unsupervised, variational approach to the inference and learning of the underlying structure of voxelized 3D shapes. Our model successfully learns 3D shapes via a hierarchical latent representation, made possible through the use of skip-connections. Realistic 3D objects can be generated by sampling its latent probabilistic manifold. We show that our inference and generative model can be trained end-to-end from 2D images to perform single image 3D model retrieval. Experiments show the improved performance of our model both quantitatively and qualitatively over a range of tasks.