Semi-Supervised Point Cloud Semantic Segmentation with Mean Teacher

Abstract

Point cloud semantic segmentation has become a key task of 3D vision scene understanding. It has made great progress in recent years, but existing methods rely heavily on vast labeled data which is expensive and time-consuming. Advances in semi-supervised learning (SSL) show that a semi-supervised learning paradigm can effectively improve model performance by adding unlabeled data for joint training with labeled data. Thus, we explore a novel semi-supervised 3D point cloud semantic segmentation framework by exploiting the Mean Teacher paradigm to utilize labeled and unlabeled data. Furthermore, we design two loss functions to force the teacher model and the student model to have the same prediction. The architecture is simple but effective, and extensive experiments demonstrate that our performance can be improved consistently by using semi-supervised learning with labeled and unlabeled data.