Created by Hugues THOMAS

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

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.

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.

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

Our code uses the nanoflann library.

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

• 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.