Copyright 2017 Cylance Inc.

Generalized Convolutional Neural Networks for Point Cloud Data

This repo contains the code used for experiments described in our paper Generalized Convolutional Neural Networks for Point Cloud Data, presented by Aleksandr Savchenkov during the poster session at the IEEE International Conference On Machine Learning and Applications (ICMLA 2017).

The goal of this work was to find a way to apply convolutional neural networks to point cloud data with arbitrary spatial features. By creating a mapping of nearest neighbors in a dataset, and applying a small shared neural network to the spatial relationships between points, we achieve an architecture that works directly with point clouds, but closely resembles a convolutional neural net in both design and behavior. Such a method bypasses the need for extensive feature engineering, while proving to be computationally efficient and requiring few parameters.

Getting Started

Requirements

• numpy
• TensorFlow
• trimesh (for vectorization)
• pyflann (for nearest neighbor computation)

Usage

1. Vectorize:

   ./vectorize.sh

   vectorize.sh uses the hyperparameters used in the paper.
2. Train:

   python train.py

   Between epochs 10 and 20 you should see test performance hit between 92.2% and 92.8% (there is some randomness).