/KPConv

Kernel Point Convolutions

Primary LanguagePythonMIT LicenseMIT

Intro figure

Created by Hugues THOMAS

Introduction

Update 27/04/2020: New PyTorch implementation available. With SemanticKitti, and Windows supported.

This repository contains the implementation of Kernel Point Convolution (KPConv), a point convolution operator presented in our ICCV2019 paper (arXiv). If you find our work useful in your research, please consider citing:

@article{thomas2019KPConv,
    Author = {Thomas, Hugues and Qi, Charles R. and Deschaud, Jean-Emmanuel and Marcotegui, Beatriz and Goulette, Fran{\c{c}}ois and Guibas, Leonidas J.},
    Title = {KPConv: Flexible and Deformable Convolution for Point Clouds},
    Journal = {Proceedings of the IEEE International Conference on Computer Vision},
    Year = {2019}
}

Update 03/05/2019, bug found with TF 1.13 and CUDA 10. We found an internal bug inside tf.matmul operation. It returns absurd values like 1e12, leading to the apparition of NaNs in our network. We advise to use the code with CUDA 9.0 and TF 1.12. More info in issue #15

SemanticKitti Code: You can download the code used for SemanticKitti submission here. It is not clean, has very few explanations, and and could be buggy. Use it only if you are familiar with KPConv implementation.

Installation

A step-by-step installation guide for Ubuntu 16.04 is provided in INSTALL.md. Windows is currently not supported as the code uses tensorflow custom operations.

Experiments

We provide scripts for many experiments. The instructions to run these experiments are in the doc folder.

  • Object Classification: Instructions to train KP-CNN on an object classification task (Modelnet40).

  • Object Segmentation: Instructions to train KP-FCNN on an object segmentation task (ShapeNetPart)

  • Scene Segmentation: Instructions to train KP-FCNN on several scene segmentation tasks (S3DIS, Scannet, Semantic3D, NPM3D).

  • New Dataset: Instructions to train KPConv networks on your own data.

  • Pretrained models: We provide pretrained weights and instructions to load them.

  • Visualization scripts: Instructions to use the three scripts allowing to visualize: the learned features, the kernel deformations and the Effective Receptive Fields.

Performances

The following tables report the current performances on different tasks and datasets. Some scores have been improved since the article submission.

Classification and segmentation of 3D shapes

Method ModelNet40 OA ShapeNetPart classes mIoU ShapeNetPart instances mIoU
KPConv rigid 92.9% 85.0% 86.2%
KPConv deform 92.7% 85.1% 86.4%

Segmentation of 3D scenes

Method Scannet mIoU Sem3D mIoU S3DIS mIoU NPM3D mIoU
KPConv rigid 68.6% 74.6% 65.4% 72.3%
KPConv deform 68.4% 73.1% 67.1% 82.0%

Acknowledgment

Our code uses the nanoflann library.

License

Our code is released under MIT License (see LICENSE file for details).

Updates

  • 17/02/2020: Added a link to SemanticKitti code
  • 24/01/2020: Bug fixes
  • 01/10/2019: Adding visualization scripts.
  • 23/09/2019: Adding pretrained models for NPM3D and S3DIS datasets.
  • 03/05/2019: Bug found with TF 1.13 and CUDA 10.
  • 19/04/2019: Initial release.