PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation
Created by Charles R. Qi, Hao Su, Kaichun Mo, Leonidas J. Guibas from Stanford University.
Introduction
This work is based on our arXiv tech report, which is going to appear in CVPR 2017. We proposed a novel deep net architecture for point clouds (as unordered point sets). You can also check our project webpage for a deeper introduction.
Point cloud is an important type of geometric data structure. Due to its irregular format, most researchers transform such data to regular 3D voxel grids or collections of images. This, however, renders data unnecessarily voluminous and causes issues. In this paper, we design a novel type of neural network that directly consumes point clouds, which well respects the permutation invariance of points in the input. Our network, named PointNet, provides a unified architecture for applications ranging from object classification, part segmentation, to scene semantic parsing. Though simple, PointNet is highly efficient and effective.
In this repository, we release code and data for training a PointNet classification network on point clouds sampled from 3D shapes, as well as for training a part segmentation network on ShapeNet Part dataset.
Citation
If you find our work useful in your research, please consider citing:
@article{qi2016pointnet,
title={PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation},
author={Qi, Charles R and Su, Hao and Mo, Kaichun and Guibas, Leonidas J},
journal={arXiv preprint arXiv:1612.00593},
year={2016}
}
Installation
Install TensorFlow. You may also need to install h5py. The code has been tested with Python 2.7, TensorFlow 1.0.1, CUDA 8.0 and cuDNN 5.1 on Ubuntu 14.04.
If you are using PyTorch, you can find a third-party pytorch implementation here.
To install h5py for Python:
sudo apt-get install libhdf5-dev
sudo pip install h5pyUsage
To train a model to classify point clouds sampled from 3D shapes:
python train.py
Log files and network parameters will be saved to log folder in default. Point clouds of ModelNet40 models in HDF5 files will be automatically downloaded (416MB) to the data folder. Each point cloud contains 2048 points uniformly sampled from a shape surface. Each cloud is zero-mean and normalized into an unit sphere. There are also text files in data/modelnet40_ply_hdf5_2048 specifying the ids of shapes in h5 files.
To see HELP for the training script:
python train.py -h
We can use TensorBoard to view the network architecture and monitor the training progress.
tensorboard --logdir log
After the above training, we can evaluate the model and output some visualizations of the error cases.
python evaluate.py --visu
Point clouds that are wrongly classified will be saved to dump folder in default. We visualize the point cloud by rendering it into three-view images.
If you'd like to prepare your own data, you can refer to some helper functions in utils/data_prep_util.py for saving and loading HDF5 files.
Part Segmentation
To train a model for object part segmentation, firstly download the data:
cd part_seg
sh download_data.sh
The downloading script will download ShapeNetPart dataset (around 1.08GB) and our prepared HDF5 files (around 346MB).
Then you can run train.py and test.py in the part_seg folder for training and testing (computing mIoU for evaluation).
License
Our code is released under MIT License (see LICENSE file for details).
TODO
Add test script for evaluation on OOS shape or point cloud data.