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

2017/03/28 by Maxim Tatarchenko, Alexey Dosovitskiy, Tatarchenko, Maxim +3 · 6 citations
Computer Science · Engineering · #3D Shape Modeling and Analysis #Advanced Vision and Imaging #Computer Graphics and Visualization Techniques #Computer Vision and Pattern Recognition (cs.CV) #FOS: Computer and information sciences

paper · pdf · doi:10.48550/arxiv.1703.09438

openalex publication_date 2017/03/28 · openalex created_date 2022/10/02 · openalex updated_date 2026/07/28

Abstract

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

Citations

• ShapeNet: An Information-Rich 3D Model Repository
• Unsupervised Learning of 3D Structure from Images
• Caffe: Convolutional Architecture for Fast Feature Embedding
• A Point Set Generation Network for 3D Object Reconstruction from a Single Image
• 3D Shape Induction from 2D Views of Multiple Objects
• 3D U-Net: Learning Dense Volumetric Segmentation from Sparse Annotation
• Learning Deconvolution Network for Semantic Segmentation

Cited by

• Evaluating Latent Generative Paradigms for High-Fidelity 3D Shape Completion from a Single Depth Image
• A 3D Generation Framework from Cross Modality to Parameterized Primitive
• StrucADT: Generating Structure-controlled 3D Point Clouds with Adjacency Diffusion Transformer
• TriLift: Interpolation-Free Tri-Plane Lifting for Efficient 3D Perception on Embedded Systems
• Multispectral-NeRF:a multispectral modeling approach based on neural radiance fields
• RayletDF: Raylet Distance Fields for Generalizable 3D Surface Reconstruction from Point Clouds or Gaussians

Related

• RocNet: Recursive Octree Network for Efficient 3D Deep Representation
• OctNet: Learning Deep 3D Representations at High Resolutions
• O-CNN
• ROAD: Learning an Implicit Recursive Octree Auto-Decoder to Efficiently Encode 3D Shapes
• Deep Octree-based CNNs with Output-Guided Skip Connections for 3D Shape and Scene Completion