Large Scale Road Datasets and Point-Offset Network for 3D Instance Segmentation

Abstract

In the field of autonomous driving, recognition and segmentation of road point clouds is an important task for the automatic production of 3D high-precision maps. To address the problems of lack of large-scale and complex road scene datasets for the instance segmentation, and the poor applicability of algorithms under large scenes, this paper produces a brand new and large-scale road instance segmentation dataset. Meanwhile, this paper proposes a brand new solution for semantic segmentation and clustering bias prediction, based on an improved Pointnet++ network, which is used together with the clustering algorithm of DBSCAN to conduct the instance segmentation. Thorough experiments indicate that the semantic segmentation accuracy of the proposed method reaches 0.982 on our produced road instance segmentation datasets, meanwhile the average accuracy and recall of the three classes of instance segmentation reach 0.853 and 0.784, respectively. Moreover, the bias network branch proposed in this paper can further improve the effectiveness of clustering, and the precision of our algorithm was improved by 15.1% and the recall rate was improved by 16.2%. It can be concluded that our produced dataset can support the large-scale road instance segmentation and our proposed algorithm can better adapt to the instance segmentation under large-scale and complex road scenarios.