Image-to-Voxel Model Translation with Conditional Adversarial Networks

We present a single-view voxel model prediction method that uses generative adversarial networks. Our method utilizes correspondences between 2D silhouettes and slices of a camera frustum to predict a voxel model of a scene with multiple object instances. We exploit pyramid shaped voxel and a generator network with skip connections between 2D and 3D feature maps. We collected two datasets VoxelCity and VoxelHome to train our framework with 36,416 images of 28 scenes with ground-truth 3D models, depth maps, and 6D object poses. We made the datasets publicly available (http://www.zefirus.org/Z_GAN). We evaluate our framework on 3D shape datasets to show that it delivers robust 3D scene reconstruction results that compete with and surpass state-of-the-art in a scene reconstruction with multiple non-rigid objects.