Point cloud classification based on transformer. (December 2022)

• Record Type: Journal Article
• Authors: Wu, Xianfeng; Liu, Xinyi; Wang, Junfei; Lai, Zhongyuan; Zhou, Jing; Liu, Xia
• Is Part Of: Computers & electrical engineering. Volume 104:Part A(2022)
• Publication Date: 2022-12
• Subjects: Point cloud classification -- PointNet -- Transformer -- Classification accuracy -- Stability; Computer engineering -- Periodicals; Electrical engineering -- Periodicals; Electrical engineering -- Data processing -- Periodicals; Ordinateurs -- Conception et construction -- Périodiques; Électrotechnique -- Périodiques; Électrotechnique -- Informatique -- Périodiques; Computer engineering; Electrical engineering; Electrical engineering -- Data processing; Periodicals; Electronic journals; 621.302854
• DOI: 10.1016/j.compeleceng.2022.108413
• Languages: English
• ISSNs: 0045-7906

Abstract

Highlights: We solve the problem that it is difficult to balance the accuracy and stability of point cloud classification. We combine a stable backbone, i.e., PointNet, with three accuracy enhancement modules, i.e., three transformers. Our method achieves a high accuracy with moderate memory and low processor consumptions. Our method is extremely stable even when the number of points is drastically reduced. Abstract: PointNet is a deep neural network that directly takes 3D point cloud data as inputs. Due to its strong stability and computational efficiency, PointNet has become one of the most popular point cloud classification methods in the real applications. Recently, transformer techniques have achieved great successes in classifying objects with image inputs, which inspires us to transplant transformers into the object classification with point cloud inputs. In this paper, we propose a 3D point cloud classification method based on PointNet with transformers. Firstly, an offset-attention module is added after the spatial alignment network and multi-layer perceptron (MLP) of PointNet. Then, a cascade-attention module (CA) is introduced into the feature alignment network of PointNet. Finally, a CA is added after the feature alignment network and MLP of PointNet. Experimental results show that our method outperforms PointNet in terms of not only classification accuracy but also stability. Graphical abstract: Image, graphical abstract