This paper introduces Occupancy Networks, a novel representation for learning-based 3D reconstruction. Unlike traditional methods that use voxel-based, point-based, or mesh-based representations, Occupancy Networks implicitly represent 3D surfaces as the continuous decision boundary of a deep neural network classifier. This approach allows for infinite resolution without excessive memory usage. The authors validate the effectiveness of their method through experiments on various input types, including single images, noisy point clouds, and coarse voxel grids. They demonstrate that their method can generate high-quality 3D meshes and outperforms state-of-the-art baselines in terms of both qualitative and quantitative metrics. The paper also includes a detailed analysis of the representation power of Occupancy Networks and an ablation study to understand the impact of different components on performance. Overall, the authors believe that Occupancy Networks will be a valuable tool for a wide range of learning-based 3D tasks.This paper introduces Occupancy Networks, a novel representation for learning-based 3D reconstruction. Unlike traditional methods that use voxel-based, point-based, or mesh-based representations, Occupancy Networks implicitly represent 3D surfaces as the continuous decision boundary of a deep neural network classifier. This approach allows for infinite resolution without excessive memory usage. The authors validate the effectiveness of their method through experiments on various input types, including single images, noisy point clouds, and coarse voxel grids. They demonstrate that their method can generate high-quality 3D meshes and outperforms state-of-the-art baselines in terms of both qualitative and quantitative metrics. The paper also includes a detailed analysis of the representation power of Occupancy Networks and an ablation study to understand the impact of different components on performance. Overall, the authors believe that Occupancy Networks will be a valuable tool for a wide range of learning-based 3D tasks.