• PyTorch implementation of PointNet++ based on erikwijmans/Pointnet2_PyTorch.
• Faster than the original codes by re-implementing the CUDA operations.

• Linux (tested on Ubuntu 14.04/16.04)
• Python 3.6+
• PyTorch 1.0

Install this library by running the following command:

cd pointnet2
python setup.py install
cd ../

Here I provide a simple example to use this library in the task of KITTI ourdoor foreground point cloud segmentation, and you could refer to the paper PointRCNN for the details of task description and foreground label generation.

1. Download the training data from KITTI 3D object detection website and organize the downloaded files as follows:

Pointnet2.PyTorch
├── pointnet2
├── tools
│   ├──data
│   │  ├── KITTI
│   │  │   ├── ImageSets
│   │  │   ├── object
│   │  │   │   ├──training
│   │  │   │      ├──calib & velodyne & label_2 & image_2
│   │  train_and_eval.py

2. Run the following command to train and evaluate:

cd tools
python train_and_eval.py --batch_size 8 --epochs 100 --ckpt_save_interval 2 

• PointRCNN: 3D object detector from raw point cloud.

• charlesq34/pointnet2: Paper author and official code repo.
• erikwijmans/Pointnet2_PyTorch: Initial work of PyTorch implementation of PointNet++.