AB-PointNet for 3D Point Cloud Recognition

Abstract

Semantic segmentation of 3D point clouds is difficult task due to its unordered representation. PointNet is a pioneering work which used 3D point clouds directly to predict 3D point semantic labels. However, it has a problem that it predicts labels without using local structure in metric space. Recent researches tackled this problem and achieved better performance. In addition to the problem, we considered that treating all channels with the same weight is obstacle to improve the accuracy. Therefore, we propose AB-PointNet which has been modified to predict 3D point semantic labels by considering the importance of channels. To emphasize the important channels, we used attention module which emphasizes channels that are useful for prediction and suppresses unimportant channels. This makes it possible to learn more effective features. In experiments, we evaluate our method on the large-scale indoor spaces 3D point cloud dataset with 13 semantic labels. Our proposed AB-PointNet has advanced performance of 3.2% in mean IoU in comparison with the conventional PointNet.