融合几何注意力与信息选择的点云处理模型

doi:10.3778/j.issn.1002-8331.2403-0293

摘要/Abstract

摘要： 现有的针对点云数据的自注意力机制虽然尝试使用流形上的测地线距离处理数据，但通常缺乏有效的预处理，导致数据无法充分适应流形特性。同时，传统的几何注意力机制存在过度聚焦的问题，无法在数据的不同重要部分之间合理分配注意力。为了提高点云数据表示的有效性并增强模型的鲁棒性，提出了一种基于几何注意力机制和信息选择的点云分类与分割模型。该模型采用类似PointNet的结构以增强特征提取能力和网络性能。在计算测地线距离之前，通过立体投影预处理数据，使其更贴合流形特征分布，从而改进了几何注意力机制的有效性。引入一种结合KL散度和熵的信息选择方法，以平衡注意力分布并提取更具代表性的特征，解决了几何注意力机制不稳定的问题。实验结果表明，与当前主流的高性能预训练模型相比，该模型在ModelNet40和ScanObjectNN数据集上的分类精度分别达到了94.9%和87.9%，整体性能提高了0.1~2.1个百分点。此外，在ShapeNetPart数据集的分割实验中，该模型同样展示了稳定的性能，凸显了信息选择模块在处理复杂点云数据中的有效性。

关键词: 深度学习, 点云分类, 点云分割, 注意力机制, 几何距离, 测地线, 信息选择模块

Abstract: Existing self-attention mechanisms for point cloud data utilize geodesic distances on manifolds but often lack effective preprocessing, hindering adaptation to manifold characteristics. Furthermore, traditional geometric attention mechanisms suffer from over-focusing, failing to distribute attention adequately across critical data regions. To enhance point cloud data representation and increase model robustness, a classification and segmentation model based on geometric attention mechanisms and information selection is proposed. The model employs a PointNet-like architecture to boost feature extraction capabilities and network performance. Data is preprocessed through stereographic projection before computing geodesic distances, allowing it to better conform to manifold characteristics, thereby improving the effectiveness of the geometric attention mechanism. Then, an information selection method combining KL divergence and entropy is introduced to balance the attention distribution and extract more representative features, addressing the instability problem of traditional geometric attention mechanisms. Experimental results demonstrate that the proposed model achieves classification accuracies of 94.9% on ModelNet40 and 87.9% on ScanObjectNN, with performance improvements ranging from 0.1 to 2.1 percentage points. In segmentation tasks on the ShapeNetPart dataset, the model also shows stable performance, highlighting the effectiveness of the information selection module in processing complex point cloud data.

Key words: deep learning, point cloud classification, point cloud segmentation, attention mechanism, geometric distance, geodesic, information selection module

刘腊梅, 柳志强. 融合几何注意力与信息选择的点云处理模型[J]. 计算机工程与应用, 2025, 61(12): 267-278.

LIU Lamei, LIU Zhiqiang. Point Cloud Processing Model Integrating Geometric Attention and Information Selection[J]. Computer Engineering and Applications, 2025, 61(12): 267-278.

参考文献

[1] 任柯燕, 谷美颖, 袁正谦, 等. 自动驾驶3D目标检测研究综述[J]. 控制与决策, 2023, 38(4): 865-889.
REN K Y, GU M Y, YUAN Z Q, et al. 3D object detection algorithms in autonomous driving: a review[J]. Control and Decision, 2023, 38(4): 865-889.
[2] 贾晓辉, 冯重阳, 刘今越. 基于点云匹配的AR饰面作业系统跟踪注册方法[J]. 计算机工程与应用, 2023, 59(6): 291-298.
JIA X H, FENG C Y, LIU J Y. Tracking and registration method based on point cloud matching for augmented reality facing work system[J]. Computer Engineering and Applications, 2023, 59(6): 291-298.
[3] 方威扬, 林东鑫, 寇万福, 等. 医学图像三维重建系统的研究进展[J]. 中国医学物理学杂志, 2022, 39(7): 823-827.
FANG W Y, LIN D X, KOU W F, et al. Advances in medical image three?dimensional reconstruction system[J]. Chinese Journal of Medical Physics, 2022, 39(7): 823-827.
[4] 杨春雨, 张鑫. 煤矿机器人环境感知与路径规划关键技术[J]. 煤炭学报, 2022, 47(7): 2844-2872.
YANG C Y, ZHANG X. Key technologies of coal mine robots for environment perception and path planning[J]. Journal of China Coal Society, 2022, 47(7): 2844-2872.
[5] 党吉圣, 杨军. 深度图注意力CNN的三维模型识别[J]. 计算机科学与探索, 2021, 15(1): 141-149.
DANG J S, YANG J. 3D model recognition based on deep graph attention CNN[J]. Journal of Frontiers of Computer Science and Technology, 2021, 15(1): 141-149.
[6] MATURANA D, SCHERER S. VoxNet: a 3D convolutional neural network for real-time object recognition[C]//Proceedings of the 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2015: 922-928.
[7] SU H, MAJI S, KALOGERAKIS E, et al. Multi-view convolutional neural networks for 3D shape recognition[C]//Proceedings of the IEEE International Conference on Computer Vision, 2015: 945-953.
[8] QI C R, SU H, MO K, et al. PointNet: deep learning on point sets for 3D classification and segmentation[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017: 652-660.
[9] GUO M H, CAI J X, LIU Z N, et al. PCT: point cloud transformer[J]. Computational Visual Media, 2021, 7(2): 187-199.
[10] LIU Y H, TIAN B, LV Y S, et al. Point cloud classification using content-based transformer via clustering in feature space[J]. IEEE/CAA Journal of Automatica Sinica, 2024, 11(1): 231-239.
[11] LI Z, TANG X, XU Z, et al. Geodesic self-attention for 3D point clouds[C]//Advances in Neural Information Processing Systems, 2022: 6190-6203.
[12] QI C R, YI L, SU H, et al. PointNet++: deep hierarchical feature learning on point sets in a metric space[C]//Advances in Neural Information Processing Systems, 2017.
[13] THOMAS H, QI C R, DESCHAUD J E, et al. KPConv: flexible and deformable convolution for point clouds[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision, 2019: 6411-6420.
[14] XU M, DING R, ZHAO H, et al. PAConv: position adaptive convolution with dynamic kernel assembling on point clouds[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021: 3173-3182.
[15] WANG Y, SUN Y B, LIU Z W, et al. Dynamic graph CNN for learning on point clouds[J]. ACM Transactions on Graphics, 2019, 38(5): 1-12.
[16] WANG L, HUANG Y, HOU Y, et al. Graph attention convolution for point cloud semantic segmentation[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019: 10296-10305.
[17] VASWANI A, SHAZEER N, PARMAR N, et al. Attention is all you need[C]//Advances in Neural Information Processing Systems, 2017.
[18] WANG X, GIRSHICK R, GUPTA A, et al. Non-local neural networks[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018: 7794-7803.
[19] ZHAO H, JIANG L, JIA J, et al. Point Transformer[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021: 16259-16268.
[20] RONNEBERGER O, FISCHER P, BROX T. U-Net: convolutional networks for biomedical image segmentation[C]//Proceedings of the International Conference on Medical Image Computing and Computer-Assisted Intervention. Cham: Springer International Publishing, 2015: 234-241.
[21] LAI X, LIU J, JIANG L, et al. Stratified Transformer for 3D point cloud segmentation[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022: 8500-8509.
[22] LIU Z, LIN Y, CAO Y, et al. Swin Transformer: hierarchical vision Transformer using shifted windows[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021: 10012-10022.
[23] ZEID K A, SCHULT J, HERMANS A, et al. Point2Vec for self-supervised representation learning on point clouds[C]//Proceedings of the 45th DAGM German Conference on Pattern Recognition. Cham: Springer Nature Switzerland, 2024: 131-146.
[24] LOSHCHILOV I, HUTTER F. Decoupled weight decay regularization[J]. arXiv:1711.05101, 2017.
[25] LOSHCHILOV I, HUTTER F. SGDR: stochastic gradient descent with warm restarts[J]. arXiv:1608.03983,2016.
[26] LI Y, BU R, SUN M, et al. PointCNN: convolution on x-transformed points[C]//Advances in Neural Information Processing Systems, 2018.
[27] YU X, TANG L, RAO Y, et al. Point-BERT: pre-training 3D point cloud Transformers with masked point modeling[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022: 19313-19322.
[28] PANG Y, WANG W, TAY F E, et al. Masked autoencoders for point cloud self-supervised learning[C]//Proceedings of the European Conference on Computer Vision, 2022: 604-621.
[29] ZHANG R, GUO Z, GAO P, et al. Point-M2AE: multi-scale masked autoencoders for hierarchical point cloud pre-training[C]//Advances in Neural Information Processing Systems, 2022: 27061-27074.