Octree Generating Networks: Efficient Convolutional Architectures for High-resolution 3D Outputs

TLDR

The paper proposes a deep convolutional decoder that generates volumetric 3D outputs efficiently using an octree representation. The architecture predicts octree structure and cell occupancies, avoiding cubic complexity by operating on an octree rather than regular voxel grids. The method enables high‑resolution 3D shape generation with limited memory, demonstrated across autoencoders, object and scene synthesis, and single‑image shape reconstruction.

Abstract

We present a deep convolutional decoder architecture that can generate volumetric 3D outputs in a compute- and memory-efficient manner by using an octree representation. The network learns to predict both the structure of the octree, and the occupancy values of individual cells. This makes it a particularly valuable technique for generating 3D shapes. In contrast to standard decoders acting on regular voxel grids, the architecture does not have cubic complexity. This allows representing much higher resolution outputs with a limited memory budget. We demonstrate this in several application domains, including 3D convolutional autoencoders, generation of objects and whole scenes from high-level representations, and shape from a single image.