Category-Augmented Net for Point Cloud Semantic Segmentation

Semantic segmentation on point cloud plays an important role in robotic applications. However, current point cloud interpolation methods have shortcomings. Distance-based interpolation schemes may confuse different semantic information, while feature-based interpolation schemes suffer from unstable training processes. To tackle this issue, we propose Category-Augmented Net (CAN). The key module of CAN is Category-Augmented Interpolation (CAI), which is used for point cloud upsampling. CAI applies Category-Augmented KNN to find the point set for interpolation in a hierarchical manner, which takes both distance and semantic information into account. Furthermore, we also introduce a dynamic training strategy, allowing the structure of our model to change during training, which makes the training progress much faster. Specifically, in the early stage of training, CAI relies on ground truth categories to select interpolation point sets, while in the later stage, CAI utilizes the predicted feature vector for this selection process. Extensive experiments on Semantic Cluttered Objects (SCO) show the effectiveness of our method.