Occupancy Networks: Learning 3D Reconstruction in Function Space

Occupancy Networks propose a new representation for learning-based 3D reconstruction, motivated by the lack in 3D of a canonical representation that is both computationally and memory efficient while allowing high-resolution geometry of arbitrary topology. Existing state-of-the-art learning-based approaches could therefore only represent coarse 3D geometry or were limited to restricted domains. The method implicitly represents the 3D surface as the continuous decision boundary of a deep neural network classifier, so the representation encodes a description of the 3D output at effectively infinite resolution without an excessive memory footprint.

The authors validate that this representation can efficiently encode 3D structure and can be inferred from various kinds of input. Their experiments demonstrate competitive results, both qualitatively and quantitatively, on the challenging tasks of 3D reconstruction from single images, noisy point clouds, and coarse discrete voxel grids. They argue that occupancy networks will become a useful tool across a wide variety of learning-based 3D tasks, and the approach became influential in implicit neural 3D representation research.