Teaching 3D to 2D generative models.
Check out Scene Representation Networks, where we replace the voxel grid with a continuous function that naturally generalizes across scenes and smoothly parameterizes scene surfaces!
Deep Generative Models today allow us to perform highly-realistic image synthesis. While each generated image is of high quality, a major challenge is to generate a series of coherent views of the same scene. This requires the network to have a latent space representation that fundamentally understands the 3D layout of the scene; e.g., how would the same chair look from a different viewpoint?
Unfortunately, this is challenging for existing models that are based on a series of 2D convolution kernels. Instead of parameterizing 3D transformations, they will explain training data in a higher-dimensional feature space, leading to poor generalization to novel views at test time - such as the output of Pix2Pix trained on images of the cube above.
With DeepVoxels, we introduce a 3D-structured neural scene representation. DeepVoxels encodes the view-dependent appearance of a 3D scene without having to explicitly model its geometry. DeepVoxels is based on a Cartesian 3D grid of persistent features that learn to make use of the underlying 3D scene structure. It combines insights from 3D computer vision with recent advances in learning image-to-image mappings. DeepVoxels is supervised, without requiring a 3D reconstruction of the scene, using a 2D re-rendering loss and enforces perspective and multi-view geometry in a principled manner.