LAF-Net: Local attention fusion for point cloud semantic segmentation

Abstract

How to utilize locally implied geometric features for points has attracted more and more attention in recent past. To tackle this dilemma, we present a novel local attention fusion module for 3D points semantic segmentation, called LAF-Net, which fuses low-dimensional contents and high-dimensional semantic features to get multi-resolutional features for points. With a modest computation cast, our LAF-Net achieves better experimental results than the several methods.