Faster Dynamic Graph CNN: Faster Deep Learning on 3D Point Cloud Data

Jinseok Hong,Hongchul Lee,Keeyoung Kim

doi:10.1109/access.2020.3023423

Jinseok Hong, Hongchul Lee + Show 1 more

Open Access

https://doi.org/10.1109/access.2020.3023423

Copy DOI

Abstract

Geometric data are commonly expressed using point clouds, with most 3D data collection devices outputting data in this form. Research on processing point cloud data for deep learning is ongoing. However, it has been difficult to apply such data as input to a convolutional neural network (CNN) or recurrent neural network (RNN) because of their unstructured and unordered features. In this study, this problem was resolved by arranging point cloud data in a canonical space through a graph CNN. The proposed graph CNN works dynamically at each layer of the network and learns the global geometric features by capturing the neighbor information of the points. In addition, by using a squeeze-and-excitation module that recalibrates the information for each layer, we achieved a good trade-off between the performance and the computation cost, and a residual-type skip connection network was designed to train the deep models efficiently. Using the proposed model, we achieved a state-of-the-art performance in terms of classification and segmentation on benchmark datasets, namely ModelNet40 and ShapeNet, while being able to train our model 2 to 2.5 times faster than other similar models.

Highlights

The point cloud is the simplest form in which data can be expressed
We propose a deep neural networks (DNNs) with 3D point cloud data as the input
The dynamic graph convolutional neural network (DGCNN) model is constructed based on a multilayer perceptron (MLP), whereas we built a deeper and faster network by adding our own skip-connection network and recalibration block

Summary

INTRODUCTION

The point cloud is the simplest form in which data can be expressed. Advances in technologies, such as Light Detection and Ranging (LIDAR) and three-dimensional (3D) scanning, have enabled acquiring 3D point cloud forms quickly. In PointNet [9], [10] a CNN model was used with point cloud input, bringing significant research attention to geometric deepening-related algorithms [49] that use non-Euclidean data, such as graphics and manifolds. The SE operation is applied to each feature map through an edge convolution block and combined to create a point cloud feature. A deeper feature map can be constructed by adding a channel-specific weighted SE operation output at each step Through this process, high-dimensional point cloud data can be processed more efficiently, and a high learning speed and improvement in performance can be expected with negligible additional computation. The model predicts an n × p label for segmentation

EXPERIMENT

Findings

CONCLUSION