This repository is to implement PointNet using PyTorch DL library, which is a deep learning network architecture proposed in 2016 by Stanford researchers and is the first neural network to handle directly 3D point clouds.

Alex Choi

• PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation
• Spatial Transformer Networks
• Medium Article on PointNet

git clone https://github.com/gchoi/PointNet.git

cd PointNet
pip -install -r requirements.txt

You can download the ModelNet dataset from here.

This code has been tested only for ModelNet-10 dataset.

Or you can download all-in-one dataset from this link.

Once you've finished downloading the dataset, please set the folder structure as follows.

• PointNet root path
  • Data
    • ModelNet10
    • ModelNet40

Of course, you can choose either one dataset of them.

This repository provide four main Python scripts:

• trainer.py which performs PointNet training.
• batch-data-tester.py which performs testing for the trained PointNet model from the test folder.
• single-data-tester.py which performs testing for a single point cloud data.
• sample-data-visualizer.py which performs the 3D visualizations of the sample point cloud data.

All configurations are defined in Config/configs.yaml.

After training has been done, you should be able to see the resultant folders:

• runs: Created by Tensorboard and can open it using the command line, $ tensorboard --logdir runs
• Outputs: Folders created followed by datetime of training started, which have Figures and Models.
  • Figures: Figures of confusion matrix as results of testing the trained network.
  • Models: Trained PointNet model files which has best-model.pth and epoch-{####}.pth.

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