Improving deep learning on point cloud by maximizing mutual information across layers. (November 2022)

• Record Type: Journal Article
• Authors: Wang, Di; Tang, Lulu; Wang, Xu; Luo, Luqing; Yang, Zhi-Xin
• Is Part Of: Pattern recognition. Volume 131(2022)
• Publication Date: 2022-11
• Subjects: Deep learning -- 3D vision -- Point clouds -- Mutual information; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2022.108892
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: A novel MMI-based method is proposed for improving learning the point cloud representation without additional data augmentation. The proposed maximizing mutual information (MMI) module can be flexibly applied in various point cloud networks. A novel MI-based loss function considers both local and global semantic information of hierarchical point cloud features. Experimental results demonstrate the efficacy and board applicability of our method. Abstract: It is a fundamental and vital task to enhance the perception capability of the point cloud learning network in 3D machine vision applications. Most existing methods utilize feature fusion and geometric transformation to improve point cloud learning without paying enough attention to mining further intrinsic information across multiple network layers. Motivated to improve consistency between hierarchical features and strengthen the perception capability of the point cloud network, we propose exploring whether maximizing the mutual information (MI) across shallow and deep layers is beneficial to improve representation learning on point clouds. A novel design of Maximizing Mutual Information (MMI) Module is proposed, which assists the training process of the main network to capture discriminative features of the input point clouds. Specifically, the MMI-based loss function is employed to constrain the differences of semantic information in two hierarchical features extracted from the shallow and deep layers of the network. Extensive experiments show that our method is generally applicable to point cloud tasks, including classification, shape retrieval, indoor scene segmentation, 3D object detection, and completion, and illustrate the efficacy of our proposed method and its advantages over existing ones. Our source code is available at https://github.com/wendydidi/MMI.git .