Pytorch PointNet

A Pytorch implementation of the PointNet network.

Reference: "PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation", Qi et al

Getting started

Setting up

Install the dependencies using Conda:

conda create --name pytorch_pointnet --file spec-file.txt

Available datasets

• shapenet: The dataset can be downloaded here. For this dataset, classification and segmentation (part segmentation) tasks are available.

  • Classification: Pretrained model

    

  • Segmentation: Pretrained model

    

    TODO: change the metric to mIoU for segmentation

• mnist: If not found, the dataset will be downloaded automatically. Only the classification task is avialable.

  • Classification: with only 50 points per cloud, Pretrained model

    

Training

Use the following script for training

python train.py dataset dataset_folder task output_folder 
   --number_of_points 2500
   --batch_size 32
   --epochs 50
   --learning_rate 0.001
   --number_of_workers 4
   --model_checkpoint

where:

• dataset: is one of the available datasets (e.g. shapenet)
• dataset_folder: is the path to the root dataset folder
• task: is either classification or segmentation
• output_folder: is the output_folder path where the training logs and model checkpoints will be stored
• number_of_points: is the amount of points per cloud
• batch_size: is the batch size
• epochs: is the number of training epochs
• learning_rate: is the optimizer learning rate
• number_of_workers: is the number of workers used by the data loader
• model_checkpoint: is the path to a checkpoint that is loaded before the training begins.

Infer

Use the following script for inference:

python infer.py dataset model_checkpoint point_cloud_file task 

where:

• dataset: is one of the available datasets (e.g. shapenet)
• model_checkpoint: is the path to a checkpoint that is loaded before the inference begins.
• point_cloud_file: is the path to the point cloud file to run the inference on.
• task: is either classification or segmentation

This will also output a 3d visualization of the point cloud.