sitzikbs / 3DmFV-Net

Point cloud classification in Real-time using 3DmFV representation and CNNs

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

3DmFV : Three-Dimensional Point Cloud Classification in Real-Time using Convolutional Neural Networks

Created by Yizhak Ben-Shabat (Itzik), Michael Lindenbaum, and Anath Fischer from Technion, I.I.T.

3DmFV Architecture

Introduction

This is the code for training a point cloud classification network using 3D modified Fisher Vectors.

This work will be presented in IROS 2018 in Madrid, Spain and will also be published in Robotics and Automation Letters.

Modern robotic systems are often equipped with a direct 3D data acquisition device, e.g. LiDAR, which provides a rich 3D point cloud representation of the surroundings. This representation is commonly used for obstacle avoidance and mapping. Here, we propose a new approach for using point clouds for another critical robotic capability, semantic understanding of the environment (i.e. object classification). Convolutional neural networks (CNN), that perform extremely well for object classification in 2D images, are not easily extendible to 3D point clouds analysis. It is not straightforward due to point clouds' irregular format and a varying number of points. The common solution of transforming the point cloud data into a 3D voxel grid needs to address severe accuracy vs memory size tradeoffs. In this paper we propose a novel, intuitively interpretable, 3D point cloud representation called 3D Modified Fisher Vectors (3DmFV). Our representation is hybrid as it combines a coarse discrete grid structure with continuous generalized Fisher vectors. Using the grid enables us to design a new CNN architecture for real-time point cloud classification. In a series of performance analysis experiments, we demonstrate competitive results or even better than state-of-the-art on challenging benchmark datasets while maintaining robustness to various data corruptions.

Citation

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

@article{ben20183dmfv,
  title={3DmFV: Three-Dimensional Point Cloud Classification in Real-Time Using Convolutional Neural Networks},
  author={Ben-Shabat, Yizhak and Lindenbaum, Michael and Fischer, Anath},
  journal={IEEE Robotics and Automation Letters},
  volume={3},
  number={4},
  pages={3145--3152},
  year={2018},
  publisher={IEEE}
}

Preprint:

@article{ben20173d,
  title={3D Point Cloud Classification and Segmentation using 3D Modified Fisher Vector Representation for Convolutional Neural Networks},
  author={Ben-Shabat, Yizhak and Lindenbaum, Michael and Fischer, Anath},
  journal={arXiv preprint arXiv:1711.08241},
  year={2017}
}

Installation

Install Tensorflow. You will also need to install h5py, scikit-learn.

The code was tested with Python 2.7, TensorFlow 1.2.1, CUDA 8.0.61 and cuDNN 5105 on Ubuntu 16.04. The code was reported by users to have also been tested on Windows 10 with python 3.6, tensorflow 1.12, cuda 9.0.176 and cudnn 9.0.

This code uses the infrastructure of PointNet as a template, however, many substantial changes have been made to the CNN model and point cloud representation.

Download the ModelNet40 data in this link.

Classification

Train the point cloud classification model using the default settings on ModelNet40, using:

python train_cls.py

Alternatively, you can tweak the different GMM parameters (e.g. number of gaussians ) or learning parameters (e.g. learning rate) using

python train_cls.py  --gpu=0 --log_dir='log' --batch_size=64 --num_point=1024 --num_gaussians=8 --gmm_variance=0.0156 
--gmm_type='grid' --learning_rate=0.001  --model='voxnet_pfv' --max_epoch=200 --momentum=0.9 --optimizer='adam'
 --decay_step=200000  --weight_decay=0.0 --decay_rate=0.7

The model will be saved to log directory. Consecutive runs with the same directory name will be saved in numbered subdirectories in order to prevent accidental overwrite of trained models.

License

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

Disclaimer

I am a mechanical engineer, not a software engineer. git is relatively new to me. Therefore, if you find any place I have made an error or have an improvement recommendation, I would appreciate your advice.

About

Point cloud classification in Real-time using 3DmFV representation and CNNs

License:Other


Languages

Language:Python 89.7%Language:Cuda 6.0%Language:C++ 3.8%Language:Shell 0.5%