三维点云双重补全网络

doi:10.3778/j.issn.1002-8331.2311-0299

摘要/Abstract

摘要： 受到传感器性能与采集遮挡等因素的影响，点云采集形成的连续缺失会严重影响被采集对象的形态与结构的表达。因此，点云的补全成为三维点云重建中非常重要的环节。目前经过编码器-解码器（encoder-decoder）框架来提取全局特征预测完整点云的点云补全网络，会使输入点云的几何结构被破坏，造成位移损失，使点云中点的分布不均匀。为了改善上述问题，搭建了一个点云双重补全网络，第一重生成粗糙点云，再经过第二重精细化生成更为精细的点云数据。并增加了一种基于多头自注意的全局特征感知模块和特征扩展模块。通过注意力机制感知聚合点云特征，增强每个点的特征维度，有效地增强了点之间的相关性。最终，在编码器-解码器框架中利用这两个模块，来增强点云补全的完整性和精细性，可以有效地恢复点云的结构。经过实验测试表明：在PCN数据集和Completion3D数据集中平均CD分别达到[5.98×10-3]和[6.11×10-3]，在可视化结果中与其他方法相比得到了更好的效果。

关键词: 点云补全, 多头注意力, 特征扩展, 双重网络

Abstract: Due to factors such as sensor performance and acquisition obstructions, continuous gaps in point cloud collection can significantly impact the representation of the form and structure of the captured object. Therefore, completing the point cloud becomes an essential task in three-dimensional point cloud analysis. Currently, employing an encoder-decoder framework to extract global features and predict complete point clouds in point cloud completion networks can disrupt the geometric structure of input point clouds, leading to displacement loss and uneven distribution of points within the cloud. To address these issues, a dual-stage point cloud completion network has been developed. The first stage generates coarse point clouds, followed by the refinement stage that produces finer point clouds. This approach incorporates a global feature perception module based on multi-head self-attention and a feature expansion module. By leveraging attention mechanisms to aggregate point cloud features, it enhances the feature dimensions of each point, effectively improving the inter-point correlations. Subsequently, within the encoder-decoder framework, these modules are utilized to enhance both the completeness and refinement of point cloud completion, effectively restoring the structure of the point cloud. Experimental tests have demonstrated that on the PCN dataset and Completion3D dataset, the average Chamfer distance achieves [5.98×10-3]and [6.11×10-3], respectively. The visual results show superior performance compared to other methods in the visualization results.

Key words: point cloud completion, multi-head attention, feature expansion, dual-network

吴萌, 严瑞祺, 孙增国, 赵怀栋, 何芊平. 三维点云双重补全网络[J]. 计算机工程与应用, 2025, 61(7): 297-305.

WU Meng, YAN Ruiqi, SUN Zengguo, ZHAO Huaidong, HE Qianping. 3D Point Cloud Dual Completion Network[J]. Computer Engineering and Applications, 2025, 61(7): 297-305.

参考文献

[1] BERGER M, TAGLIASACCHI A, SEVERSKY L M, et al. State of the art in surface reconstruction from point clouds[J]. Eurographics, 2014, 1: 161-185.
[2] DAVIS J, MARSCHNER S R, GARR M, et al. Filling holes in complex surfaces using volumetric diffusion[C]//Proceedings of the 1st International Symposium on 3D Data Processing Visualization and Transmission. Piscataway: IEEE, 2002: 428-441.
[3] MITRA N J, GUIBAS L J, PAULY M, et al. Partial and approximate symmetry detection for 3D geometry[J]. ACM Transactions on Graphics, 2006, 25(3): 560-568.
[4] MITRA N J, PAULY M, WAND M, et al. Symmetry in 3D geometry: extraction and applications[J]. Computer Graphics Forum, 2013, 32(6): 1-23.
[5] DAI A, QI C R, NIE?NER M. Shape completion using 3D-encoder-predictor CNNs and shape synthesis[C]//Proceedings of the 2017 IEEE Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE, 2017: 5868-5877.
[6] WANG W Y, HUANG Q G, YOU S Y, et al. Shape inpainting using 3D generative adversarial network and recurrent convolutional networks[C]//Proceedings of the 2017 IEEE International Conference on Computer Vision. Piscataway: IEEE, 2017: 2298-2306.
[7] YANG Y Q, FENG C, SHEN Y R, et al. FoldingNet: point cloud auto-encoder via deep grid deformation[C]//Proceedings of the 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE, 2018: 206-215.
[8] CHARLES R Q, HAO S, MO K C, et al. PointNet: deep learning on point sets for 3D classification and segmentation[C]//Proceedings of the 2017 IEEE Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE, 2017: 652-660.
[9] QI C R, YI L, SU H, et al. PointNet++[C]//Proceedings of the 31st International Conference on Neural Information Processing Systems. New York: ACM, 2017: 5105-5114.
[10] LI X K, LI C S, TONG Z K, et al. Campus3D: a photogrammetry point cloud benchmark for hierarchical understanding of outdoor scene[C]//Proceedings of the 28th ACM International Conference on Multimedia. New York: ACM, 2020: 238-246.
[11] YUAN W T, KHOT T, HELD D, et al. PCN: point completion network[C]//Proceedings of the 2018 International Conference on 3D Vision. Piscataway: IEEE, 2018: 728-737.
[12] TCHAPMI L P, KOSARAJU V, REZATOFIGHI H, et al. TopNet: structural point cloud decoder[C]//Proceedings of the 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE, 2019: 383-392.
[13] HUANG Z T, YU Y K, XU J W, et al. PF-Net: point fractal network for 3D point cloud completion[C]//Proceedings of the 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE, 2020: 7662-7670.
[14] XIE H Z, YAO H X, ZHOU S C, et al. GRNet: gridding residual network for dense point cloud completion[C]//Proceedings of the European Conference on Computer Vision. Cham: Springer, 2020: 365-381.
[15] WANG X G, ANG M H, LEE G H. Cascaded refinement network for point cloud completion[C]//Proceedings of the 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE, 2020: 790-799.
[16] XIANG P, WEN X, LIU Y S, et al. Snowflake point deconvolution for point cloud completion and generation with skip-transformer[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023, 45(5): 6320-6338.
[17] ZHANG W X, YAN Q G, XIAO C X, et al. Detail preserved point cloud completion via separated feature aggregation[C]//Proceedings of the 16th European Conference on Computer Vision. New York: ACM, 2020: 512-528.
[18] PAN L, CHEN X, CAI Z, et al. Variational relational point completion network[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021: 8524-8533.
[19] WEN X, XIANG P, HAN Z Z, et al. PMP-Net: point cloud completion by learning multi-step point moving paths[C]//Proceedings of the 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE, 2021: 7439-7448.
[20] WANG X G, ANG M H, LEE G H. Voxel-based network for shape completion by leveraging edge generation[C]//Proceedings of the 2021 IEEE/CVF International Conference on Computer Vision. Piscataway: IEEE, 2021: 13169-13178.
[21] ASHISH V. Attention is all you need[C]//Advances in Neural Information Processing Systems, 2017.
[22] VINICIUS M, FLORENCIA C. Point cloud transformers applied to collider physics[J]. Machine Learning: Science and Technology, 2021, 2(3): 035027.
[23] GUO M H, CAI J X, LIU Z N, et al. PCT: point cloud transformer[J]. Computational Visual Media, 2021, 7(2): 187-199.
[24] YU X M, RAO Y M, WANG Z Y, et al. PoinTr: diverse point cloud completion with geometry-aware transformers[C]//Proceedings of the 2021 IEEE/CVF International Conference on Computer Vision. Piscataway: IEEE, 2021: 12498-12507.
[25] WANG J, CUI Y, GUO D Y, et al. PointAttN: you only need attention for point cloud completion[J]. Proceedings of the AAAI Conference on Artificial Intelligence, 2024, 38(6): 5472-5480.
[26] ZHANG W X, ZHOU H J, DONG Z, et al. Point cloud completion via skeleton-detail transformer[J]. IEEE Transactions on Visualization and Computer Graphics, 2023, 29(10): 4229-4242.
[27] PAN L, CHEW C M, LEE G H. PointAtrousGraph: deep hierarchical encoder-decoder with point atrous convolution for unorganized 3D points[C]//Proceedings of the 2020 IEEE International Conference on Robotics and Automation. Piscataway: IEEE, 2020: 1113-1120.
[28] FAN H Q, SU H, GUIBAS L. A point set generation network for 3D object reconstruction from a single image[C]//Proceedings of the 2017 IEEE Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE, 2017: 605-613.
[29] VAKALOPOULOU M, CHASSAGNON G, BUS N, et al. AtlasNet: multi-atlas non-linear deep networks for medical image segmentation[C]//Proceedings of the 21st International Conference on Medical Image Computing and Computer Assisted Intervention. Cham: Springer, 2018: 658-666.
[30] WANG L, HUANG Y C, SHAN J, et al. MSNet: multi-scale convolutional network for point cloud classification[J]. Remote Sensing, 2018, 10(4): 612.
[31] NIE Y Y, LIN Y Q, HAN X G, et al. Skeleton-bridged point completion: from global inference to local adjustment[C]//Advances in Neural Information Processing Systems, 2020, 33: 16119-16130.
[32] WEN X, LI T Y, HAN Z Z, et al. Point cloud completion by skip-attention network with hierarchical folding[C]//Proceedings of the 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE, 2020: 1939-1948.
[33] LI S S, GAO P, TAN X Y, et al. ProxyFormer: proxy alignment assisted point cloud completion with missing part sensitive transformer[C]//Proceedings of the 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE, 2023: 9466-9475.
[34] ZHU Z, CHEN H H, HE X, et al. SVDFormer: complementing point cloud via self-view augmentation and self-structure dual-generator[C]//Proceedings of the 2023 IEEE/CVF International Conference on Computer Vision. Piscataway: IEEE, 2023: 14508-14518.
[35] YANG H, CHEN Y F, SONG K Y, et al. Multiscale feature-clustering-based fully convolutional autoencoder for fast accurate visual inspection of texture surface defects[J]. IEEE Transactions on Automation Science and Engineering, 2019, 16(3): 1450-1467.
[36] WANG Y D, TAN D J, NAVAB N, et al. SoftPoolNet: shape descriptor for point cloud completion and classification[C]//Proceedings of the 16th European Conference on Computer Vision. Cham: Springer, 2020: 70-85.