Exploiting Inductive Bias in Transformer for Point Cloud Classification and Segmentation

Discovering inter-point connection for efficient high-dimensional featureextraction from point coordinate is a key challenge in processing point cloud.Most existing methods focus on designing efficient local feature extractorswhile ignoring global connection, or vice versa. In this paper, we design a newInductive Bias-aided Transformer (IBT) method to learn 3D inter-pointrelations, which considers both local and global attentions. Specifically,considering local spatial coherence, local feature learning is performedthrough Relative Position Encoding and Attentive Feature Pooling. Weincorporate the learned locality into the Transformer module. The local featureaffects value component in Transformer to modulate the relationship betweenchannels of each point, which can enhance self-attention mechanism withlocality based channel interaction. We demonstrate its superiorityexperimentally on classification and segmentation tasks. The code is availableat: https://github.com/jiamang/IBT