O-CNN

Documentation

This repository contains the pure PyTorch-based implementation of O-CNN. The code has been tested with Pytorch>=1.6.0, and Pytorch>=1.9.0 is preferred. The original implementation of O-CNN is based on C++ and CUDA and can be found here, which has received and .

O-CNN is an octree-based 3D convolutional neural network framework for 3D data. O-CNN constrains the CNN storage and computation into non-empty sparse voxels for efficiency and uses the octree data structure to organize and index these sparse voxels. Currently, this type of 3D convolution is known as Sparse Convolution in the research community.

The concept of Sparse Convolution in O-CNN is the same with SparseConvNet, MinkowskiNet, and SpConv. The key difference is that our O-CNN uses octrees to index the sparse voxels, while these works use Hash Tables. However, I believe that octrees may be the right choice for Sparse Convolution. With octrees, I can implement the Sparse Convolution with pure PyTorch. More importantly, with octrees, I can also build efficient transformers for 3D data -- OctFormer, which is extremely hard with Hash Tables.

Our O-CNN is published in SIGGRAPH 2017, SparseConvNet is published in CVPR 2018, and MinkowskiNet is published in CVPR 2019. Actually, our O-CNN was submitted to SIGGRAPH in the end of 2016 and was officially accepted in March, 2017. We just did not post our paper on Arxiv during the review process of SIGGRAPH. Therefore, the idea of constraining CNN computation into sparse non-emtpry voxels, i.e. Sparse Convolution, is first proposed by our O-CNN.

Key benefits of ocnn-pytorch

• Simplicity. The ocnn-pytorch is based on pure PyTorch, it is portable and can be installed with a simple command:pip install ocnn. Other sparse convolution frameworks heavily rely on C++ and CUDA, and it is complicated to configure the compiling environment.

• Efficiency. The ocnn-pytorch is very efficient compared with other sparse convolution frameworks. It only takes 18 hours to train the network on ScanNet for 600 epochs with 4 V100 GPUs. For reference, under the same training settings, MinkowskiNet 0.4.3 takes 60 hours and MinkowskiNet 0.5.4 takes 30 hours.

Citation

@article {Wang-2017-ocnn,
  title    = {{O-CNN}: Octree-based Convolutional Neural Networksfor {3D} Shape Analysis},
  author   = {Wang, Peng-Shuai and Liu, Yang and Guo, Yu-Xiao and Sun, Chun-Yu and Tong, Xin},
  journal  = {ACM Transactions on Graphics (SIGGRAPH)},
  volume   = {36},
  number   = {4},
  year     = {2017},
}