Encoding discriminative representation for point cloud semantic segmentation

Abstract

Abstract The main challenge for semantic segmentation based on deep learning is how to encode effective representation due to the nature of point cloud. This letter proposes a plug‐and‐play module to abstract information from local, multi‐scale context, and interdependency to make up for the deficiency. Given an intermediate feature, our module first performs a multi‐scale context layer to adaptive density variance and gets rich contextual information. Then, our module infers attention masks along two dimensions, channel and pointwise in parallel. The attention masks are performed on the input features for further cultivation. Through this module, we can refine features from different perspectives. Experiments conducted on the Stanford 3D Indoor Semantics dataset demonstrate the efficacy of the proposed method. The results show that our method can achieve state‐of‐the‐art performance.