Semi-supervised point cloud segmentation using self-training with label confidence prediction

Abstract

Point cloud segmentation is a key problem in 3D content understanding. The existing methods based on deep neural network for point cloud segmentation mainly train the network in a supervised fashion, which heavily rely on a large amount of high-quality manual-labeled training point clouds. However, it is very tedious and time-consuming to manually assign part labels for each point in point clouds. Meanwhile, a lot of unlabeled point clouds can easily be obtained from 3D scanners, Internet or reconstruction. Therefore, we introduce self-training to utilize these unlabeled point clouds. So the proposed semi-supervised point cloud segmentation method can employ both labeled point clouds and unlabeled point clouds for training. Moreover, in order to make better use of unlabeled point clouds, the adopted adversarial architecture proposes confidence discrimination of label prediction for unlabeled point clouds. Thus, pseudo labels on unlabeled point clouds with higher reliability can be picked out to participate the network training, which further improves segmentation performance. The experiments show that the proposed method can make full use of the unlabeled point clouds in training. In addition, segmentation performance improves by self-training with label confidence prediction.