Existing methods for 3D point cloud semantic segmentation are not robust enough for dynamic testing environments with emerging new classes because they require training all classes at once. To address this problem, class-incremental learning continually learns new categories using previous knowledge without substantial costs of memory storage and re-training. In this paper, we propose a novel approach for 3D point cloud class-incremental semantic segmentation, which aims to identify both old and novel classes using incremental learning without forgetting old classes. In our framework, we propose (1) an extra branch classifier to recognize the background, which assists the main classifier; (2) a similarity weights estimation approach to re-weight the classifier’s predictions by fusing both geometric and deep semantic features and searching connections between similar points; and (3) a flexible pseudo-label generation strategy to combine the novel classes with the old classes from the previous model. The extensive experimental results on the S3DIS and ScanNet datasets demonstrate that our method achieves remarkable performance compared to existing methods and the joint training strategy.