Learning-Deep-Learning

Occupancy Networks: Learning 3D Reconstruction in Function Space

May 2023

tl;dr: Encoding the occupancy of a scene with a neural network, and can be queried at any location with arbitrary resolution.

Overall impression

In 3D there is no canonical representation which is both computationally and memory efficient yet allows for representing high-resolution geometry of arbitrary topology.

Occupancy networks implicitly represent the 3D surface as the continuous decision boundary of a deep neural network classifier. Instead of predicting a voxelized representation at a fixed resolution, the network can be evaluated at arbitrary resolution. This drastically reduces the memory footprint during training.

Key ideas

Technical details

Notes