A gradient optimization and manifold preserving based binary neural network for point cloud. (July 2023)

• Record Type: Journal Article
• Authors: Zhao, Zhi; Xu, Ke; Ma, Yanxin; Wan, Jianwei
• Is Part Of: Pattern recognition. Volume 139(2023)
• Publication Date: 2023-07
• Subjects: Binarization -- Gradient optimization -- Manifold -- Point cloud; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2023.109445
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: Gradient optimization alleviates gradient mismatch brought by forward binarization. Quantization causes manifold distortion for feature maps of point cloud samples. Manifold preserving based learnable scaling benefits representation fidelity more. Pooling correction based on manifold preserving alleviates severe feature homogeneity. Abstract: With significant progress of deep learning on 3D point cloud, the demand for deployment of point cloud neural network on the edge devices is growing. Binary neural network, a type of quantization compression method, with extreme low bit and fast inference speed, attracts more attention. It is more challenging, but has greater potentiality. Most of the researches on binary networks focus on images rather than point cloud. Considering the particularity of point cloud neural network, this paper presents a novel binarization framework, which includes two main contributions. Firstly, a gradient optimization method is proposed to overcome the shortcomings of Straight Through Estimator (STE) commonly used in the back propagation of binary network training. Secondly, based on the analysis of manifold distortion caused by the binary convolution and pooling operations, we propose an optimized scaling recovery method to restore manifold for the convoluted feature, and also, a pooling correction method to improve the pooled feature's fidelity. Manifold distortion leads to the severe feature homogeneity problem, which brings trouble in generating features with sufficient discrimination for classification and segmentation. The manifold preserving optimizations are designed to introduce minimum extra parameters to balance the accuracy with the computation and storage consumption. Experiments show that the proposed method outperforms state-of-the-art in accuracy with ignored overhead, and also has good scalability.