Occupancy Networks: Learning 3D Reconstruction in Function Space This paper introduces Occupancy Networks, a new representation for learning-based 3D reconstruction. Unlike traditional methods that discretize 3D space into voxels, points, or meshes, occupancy networks represent 3D surfaces as the continuous decision boundary of a deep neural network classifier. This approach allows for efficient encoding of high-resolution 3D geometry without excessive memory usage. The network can be trained to infer 3D structures from various input types, including images, point clouds, and low-resolution voxel grids. During inference, high-quality 3D meshes are extracted using a multi-resolution isosurface extraction algorithm. The method is validated through experiments on challenging tasks such as 3D reconstruction from single images, noisy point clouds, and coarse voxel grids. Results show that occupancy networks achieve competitive performance, both qualitatively and quantitatively, compared to state-of-the-art methods. The approach is memory-efficient and can represent complex 3D shapes at high resolution. It also enables the generation of high-quality 3D meshes with normal information, and can be applied to a wide range of 3D tasks. The paper also includes an ablation study showing that uniform sampling and a ResNet architecture with conditional batch normalization yield the best results. The method is shown to generalize well to real-world data, including the KITTI and Online Products datasets.Occupancy Networks: Learning 3D Reconstruction in Function Space This paper introduces Occupancy Networks, a new representation for learning-based 3D reconstruction. Unlike traditional methods that discretize 3D space into voxels, points, or meshes, occupancy networks represent 3D surfaces as the continuous decision boundary of a deep neural network classifier. This approach allows for efficient encoding of high-resolution 3D geometry without excessive memory usage. The network can be trained to infer 3D structures from various input types, including images, point clouds, and low-resolution voxel grids. During inference, high-quality 3D meshes are extracted using a multi-resolution isosurface extraction algorithm. The method is validated through experiments on challenging tasks such as 3D reconstruction from single images, noisy point clouds, and coarse voxel grids. Results show that occupancy networks achieve competitive performance, both qualitatively and quantitatively, compared to state-of-the-art methods. The approach is memory-efficient and can represent complex 3D shapes at high resolution. It also enables the generation of high-quality 3D meshes with normal information, and can be applied to a wide range of 3D tasks. The paper also includes an ablation study showing that uniform sampling and a ResNet architecture with conditional batch normalization yield the best results. The method is shown to generalize well to real-world data, including the KITTI and Online Products datasets.