PCN: Point Completion Network

[paper] [data] [website]

Introduction

PCN is a learning-based shape completion method which directly maps a partial point cloud to a dense, complete point cloud without any voxelization. It is based on our 3DV 2018 publication PCN: Point Completion Network. Please refer to our project website or read our paper for more details.

Citation

If you find our work useful for your research, please cite:

@inProceedings{yuan2018pcn,
  title     = {PCN: Point Completion Network},
  author    = {Yuan, Wentao and Khot, Tejas and Held, David and Mertz, Christoph and Hebert, Martial},
  booktitle = {3D Vision (3DV), 2018 International Conference on},
  year      = {2018}
}

Usage

1) Prerequisite

1. Install dependencies via pip3 install -r requirments.txt.
2. Follow this guide to install Open3D for point cloud I/O.
3. Build point cloud distance ops by running make under pc_distance. Make sure the paths in makefile are correct.
4. Download trained models from Google Drive.

This code is built using Tensorflow 1.12 with CUDA 9.0 and tested on Ubuntu 16.04 with Python 3.5.

2) Demo

Run python3 demo.py. Use --input_path option to switch between the input examples in demo_data.

3) ShapeNet Completion

1. Download ShapeNet test data in the shapenet folder on Google Drive. Specifically, this experiment requires test, test_novel, test.list and test_novel.list.
2. Run python3 test_shapenet.py. Use --model_type option to choose different model architectures. Type python3 test_shapenet.py -h for more options.

4) KITTI Completion

1. Download KITTI data in the kitti folder on Google Drive.
2. Run python3 test_kitti.py. Type python3 test_kitti.py -h for more options.

5) KITTI Registration

1. Run the KITTI completion experiment first to get complete point clouds.
2. Run python3 kitti_registration.py. Type python3 kitti_registration.py -h for more options.

6) Training

1. Download training (train.lmdb, train.lmdb-lock) and validation (valid.lmdb, valid.lmdb-lock) data from shapenet or shapenet_car directory on Google Drive. Note that the training data for all 8 categories in shapenet takes up 49G of disk space. The training data for only the car category takes 9G instead.
2. Run python3 train.py. Type python3 train.py -h for more options.

7) Data Generation

To generate your own data from ShapeNet, first Download ShapeNetCore.v1. Then, create partial point clouds from depth images (see instructions in render) and corresponding ground truths by sampling from CAD models (see instructions in sample). Finally, serialize the data using lmdb_writer.py.

License

This project Code is released under the MIT License (refer to the LICENSE file for details).