zlichen / o-cnn

O-CNN: Octree-based Convolutional Neural Networks for 3D Shape Analysis

O-CNN

This repository contains the implementation of O-CNN and Adaptive O-CNN introduced in our SIGGRAPH 2017 paper and SIGGRAPH Asia 2018 paper.
The code is released under the MIT license.

• O-CNN: Octree-based Convolutional Neural Networks
  By Peng-Shuai Wang, Yang Liu, Yu-Xiao Guo, Chun-Yu Sun and Xin Tong
  ACM Transactions on Graphics (SIGGRAPH), 36(4), 2017

• Adaptive O-CNN: A Patch-based Deep Representation of 3D Shapes
  By Peng-Shuai Wang, Chun-Yu Sun, Yang Liu and Xin Tong
  ACM Transactions on Graphics (SIGGRAPH Asia), 37(6), 2018
  

• Deep Octree-based CNNs with Output-Guided Skip Connections for 3D Shape and Scene Completion
  By Peng-Shuai Wang, Yang Liu and Xin Tong
  Computer Vision and Pattern Recognition (CVPR) Workshops, 2020
  

• Unsupervised 3D Learning for Shape Analysis via Multiresolution Instance Discrimination
  By Peng-Shuai Wang, Yu-Qi Yang, Qian-Fang Zou, Zhirong Wu, Yang Liu and Xin Tong
  AAAI Conference on Artificial Intelligence (AAAI), 2021. [Arxiv, 2020.08]
  

If you use our code or models, please cite our paper.

What's New?

• 2021.03.01: Update the code for pytorch-based O-CNN, including a ResNet and some important modules.
• 2021.02.08: Release the code for ShapeNet segmentation with HRNet.
• 2021.02.03: Release the code for ModelNet40 classification with HRNet.
• 2020.10.12: Release the initial version of our O-CNN under PyTorch. The code has been tested with the classification task.
• 2020.08.16: We released our code for 3D unsupervised learning. We provided a unified network architecture for generic shape analysis tasks and an unsupervised method to pretrain the network. Our method achieved state-of-the-art performance on several benchmarks.
• 2020.08.12: We released our code for Partnet segmentation. We achieved an average IoU of 58.4, significantly better than PointNet (IoU: 35.6), PointNet++ (IoU: 42.5), SpiderCNN (IoU: 37.0), and PointCNN(IoU: 46.5).
• 2020.08.05: We released our code for shape completion. We proposed a simple yet efficient network and output-guided skip connections for 3D completion, which achieved state-of-the-art performances on several benchmarks.

Contents

• Installation
• Data Preparation
• Shape Classification
• Shape Retrieval
• Shape Segmentation
• Shape Autoencoder
• Shape Completion
• Image2Shape
• Unsupverised Pretraining

We thank the authors of ModelNet, ShapeNet and Region annotation dataset for sharing their 3D model datasets with the public.

Please contact us (Pengshuai Wang [email protected], Yang Liu [email protected] ) if you have any problems about our implementation.