/pointnet2

PointNet++: Deep Hierarchical Feature Learning on Point Sets in a Metric Space

Primary LanguagePythonOtherNOASSERTION

PointNet++: Deep Hierarchical Feature Learning on Point Sets in a Metric Space

Created by Charles R. Qi, Li (Eric) Yi, Hao Su, Leonidas J. Guibas from Stanford University.

prediction example

Citation

If you find our work useful in your research, please consider citing:

    @article{qi2017pointnetplusplus,
      title={PointNet++: Deep Hierarchical Feature Learning on Point Sets in a Metric Space},
      author={Qi, Charles R and Yi, Li and Su, Hao and Guibas, Leonidas J},
      journal={arXiv preprint arXiv:1706.02413},
      year={2017}
    }

Introduction

This work is based on our NIPS'17 paper. You can find arXiv version of the paper here or check project webpage for a quick overview. PointNet++ is a follow-up project that builds on and extends PointNet. It is version 2.0 of the PointNet architecture.

PointNet (the v1 model) either transforms features of individual points independently or process global features of the entire point set. However, in many cases there are well defined distance metrics such as Euclidean distance for 3D point clouds collected by 3D sensors or geodesic distance for manifolds like isometric shape surfaces. In PointNet++ we want to respect spatial localities of those point sets. PointNet++ learns hierarchical features with increasing scales of contexts, just like that in convolutional neural networks. Besides, we also observe one challenge that is not present in convnets (with images) -- non-uniform densities in natural point clouds. To deal with those non-uniform densities, we further propose special layers that are able to intelligently aggregate information from different scales.

In this repository we release code and data for our PointNet++ classification and segmentation networks as well as a few utility scripts for training, testing and data processing and visualization.

Installation

Install TensorFlow. The code is tested under TF1.2 GPU version and Python 2.7 (version 3 should also work) on Ubuntu 14.04. There are also some dependencies for a few Python libraries for data processing and visualizations like cv2, h5py etc. It's highly recommended that you have access to GPUs.

Compile Customized TF Operators

The TF operators are included under tf_ops, you need to compile them (check tf_xxx_compile.sh under each ops subfolder) first. Update nvcc and python path if necessary. The code is tested under TF1.2.0. If you are using earlier version it's possible that you need to remove the -D_GLIBCXX_USE_CXX11_ABI=0 flag in g++ command in order to compile correctly.

To compile the operators in TF version >=1.4, you need to modify the compile scripts slightly.

First, find Tensorflow include and library paths.

    TF_INC=$(python -c 'import tensorflow as tf; print(tf.sysconfig.get_include())')
    TF_LIB=$(python -c 'import tensorflow as tf; print(tf.sysconfig.get_lib())')

Then, add flags of -I$TF_INC/external/nsync/public -L$TF_LIB -ltensorflow_framework to the g++ commands.

Usage

Shape Classification

To train a PointNet++ model to classify ModelNet40 shapes (using point clouds with XYZ coordinates):

    python train.py

To see all optional arguments for training:

    python train.py -h

If you have multiple GPUs on your machine, you can also run the multi-GPU version training (our implementation is similar to the tensorflow cifar10 tutorial):

    CUDA_VISIBLE_DEVICES=0,1 python train_multi_gpu.py --num_gpus 2

After training, to evaluate the classification accuracies (with optional multi-angle voting):

    python evaluate.py --num_votes 12

Side Note: For the XYZ+normal experiment reported in our paper: (1) 5000 points are used and (2) a further random data dropout augmentation is used during training (see commented line after augment_batch_data in train.py and (3) the model architecture is updated such that the nsample=128 in the first two set abstraction levels, which is suited for the larger point density in 5000-point samplings.

To use normal features for classification: You can get our sampled point clouds of ModelNet40 (XYZ and normal from mesh, 10k points per shape) here (1.6GB). Move the uncompressed data folder to data/modelnet40_normal_resampled

Object Part Segmentation

To train a model to segment object parts for ShapeNet models:

    cd part_seg
    python train.py

Preprocessed ShapeNetPart dataset (XYZ, normal and part labels) can be found here (674MB). Move the uncompressed data folder to data/shapenetcore_partanno_segmentation_benchmark_v0_normal

Semantic Scene Parsing

See scannet/README and scannet/train.py for details.

Visualization Tools

We have provided a handy point cloud visualization tool under utils. Run sh compile_render_balls_so.sh to compile it and then you can try the demo with python show3d_balls.py The original code is from here.

Prepare Your Own Data

You can refer to here on how to prepare your own HDF5 files for either classification or segmentation. Or you can refer to modelnet_dataset.py on how to read raw data files and prepare mini-batches from them. A more advanced way is to use TensorFlow's dataset APIs, for which you can find more documentations here.

License

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

Updates

  • 02/23/2018: Added support for multi-gpu training for the classification task.
  • 02/23/2018: Adopted a new way for data loading. No longer require manual data downloading to train a classification network.
  • 02/06/2018: Added sample training code for ScanNet semantic segmentation.

Related Projects