Accuracy Improvement of Deep Learning 3D Point Cloud Instance Segmentation

Abstract

In the field of computer vision, the research on 3D point clouds is one of the important tasks of 3D scene understanding, such as target detection, semantic segmentation, etc., which have achieved rich research results. Combining the concepts of target detection and semantic segmentation, it is not only necessary to identify point clouds of different semantics, but also to distinguish instances of the same semantics. Therefore, the research on 3D point cloud instance segmentation will be more challenging. The results of most instance segmentation methods have found that if the distances between different objects of the same semantic are too close, it is difficult to distinguish them from each other, resulting in poor accuracy of instance segmentation. In order to improve this problem, based on the method of a deep learning instance segmentation of the 3D point cloud scene, the 3D object point cloud with the same semantic meaning and the same color label is transferred in to the voxel space. The voxels of an object is then projected onto the 2D image from the top view. At last, the image segmentation for post-processing is utilized to further improve the accuracy of instance segmentation.