Skip to content

Neural implicit reconstruction experiments for the Vector Neuron paper

License

Notifications You must be signed in to change notification settings

wx-b/vnn-neural-implicits

 
 

Repository files navigation

Neural Implicit Reconstruction with Vector Neurons

This repository contains code for the neural implicit reconstruction experiments in the paper Vector Neurons: A General Framework for SO(3)-Equivariant Networks. Code for classification and segmentation experiments can be found here.

[Project] [Paper]

Preparation

The code structure follows Occupancy Networks. Please follow their instructions to prepare the data and install the dependencies. Run

python generate_random_rotation.py

to precompute the random rotations for all input pointclouds.

Usage

To train and evaluate the networks, please run these two commands

python train.py CONFIG.yaml
python eval.py CONFIG.yaml

The configuration files are, for VN-OccNet,

configs/equinet/vnn_pointnet_resnet_resnet_ROTATION.yaml

for the vanilla OccNet baseline,

configs/pointcloud/onet_resnet_ROTATION.yaml

and for vanilla PointNet encoder + invariant decoder,

configs/equinet/inner_baseline_resnet_ROTATION.yaml

ROTATION can be chosen from aligned (no rotations) and so3 (with precomputed random rotations). We also provide two settings rot-rand (generate random rotations on the fly during training) and pca (apply PCA pre-alignment the the input pointclouds), which are not reported in the paper.

Citation

Please cite this paper if you want to use it in your work,

@article{deng2021vn,
  title={Vector Neurons: a general framework for SO(3)-equivariant networks},
  author={Deng, Congyue and Litany, Or and Duan, Yueqi and Poulenard, Adrien and Tagliasacchi, Andrea and Guibas, Leonidas},
  journal={arXiv preprint arXiv:2104.12229},
  year={2021}
}

License

MIT License

Acknowledgement

The structure of this codebase is borrowed from Occupancy Networks.

About

Neural implicit reconstruction experiments for the Vector Neuron paper

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 62.7%
  • C++ 14.1%
  • Cuda 12.0%
  • C 8.3%
  • Mako 2.9%