A Convolutional Neural Network for Point Cloud Instance Segmentation in Cluttered Scene Trained by Synthetic Data Without Color

Yajun Xu,Shogo Arai,Fuyuki Tokuda,Kazuhiro Kosuge

doi:10.1109/access.2020.2978506

Yajun Xu, Shogo Arai + Show 2 more

Open Access

https://doi.org/10.1109/access.2020.2978506

Copy DOI

Abstract

3D Instance segmentation is a fundamental task in computer vision. Effective segmentation plays an important role in robotic tasks, augmented reality, autonomous driving, etc. With the ascendancy of convolutional neural networks in 2D image processing, the use of deep learning methods to segment 3D point clouds receives much attention. A great convergence of training loss often requires a large amount of human-annotated data, while making such a 3D dataset is time-consuming. This paper proposes a method for training convolutional neural networks to predict instance segmentation results using synthetic data. The proposed method is based on the SGPN framework. We replaced the original feature extractor with “dynamic graph convolutional neural networks” that learned how to extract local geometric features and proposed a simple and effective loss function, making the network more focused on hard examples. We experimentally proved that the proposed method significantly outperforms the state-of-the-art method in both Stanford 3D Indoor Semantics Dataset and our datasets.

Highlights

Segmentation is an important means to make data easier to understand and analyze
Inspired by SGPN [7], which uses a single network for performing instance segmentation on point clouds, we propose a simple and effective method
We propose a novel method for instance segmentation on point cloud without color

Summary

INTRODUCTION

Segmentation is an important means to make data easier to understand and analyze. It is helpful for robot tasks [1], autonomous driving [2], augmented reality [3], and visual servoing [4]. The RGBD image or scene point cloud contains a lot of redundant information, for instance, irrelevant objects and background. The segmentation method does mitigate computational cost and improve the precision of pose estimation [9]–[13]. Inspired by SGPN [7], which uses a single network for performing instance segmentation on point clouds, we propose a simple and effective method. The proposed method can obtain the training data by synthesis With this method, we can recognize almost all the target objects in the scene and pick out the appropriate point cloud for some robot tasks such as pose estimation and grasping.

RELATED WORK

DISTANCE-MASK

EXPERIMENT

Findings

CONCLUSIONS