Employing Offset-Attention for 3D Few-shot Semantic Segmentation

The existing deep 3D semantic segmentation methods mostly are trained with a large number of human annotations. However, due to the expensive labor for annotations label, few-shot 3D semantic segmentation is achieving more attention. In this work, we improve the performance of few-shot learning based on semantic segmentation of 3D point clouds using the offset attention method that has been successfully applied in natural language processing. Experiments demonstrate the superiority of the offset-attention in 3D semantic segmentation on the benchmark datasets S3DIS.

I. Introduction

Point cloud data, which mostly are collected from Lidar, can represent the 3D structure of the scene. Point cloud processing has been widely studied for 3D scene understanding, and has been used in many fields, including autonomous driving, virtual enhancement, and reality enhancement. Point cloud semantic segmentation is one of the basic point cloud processing tasks, which aims to predict the label of each point. However, the unordered and unstructured characteristics of point clouds make the semantic segmentation of point clouds a challenging task.