毕晓华, 牛春博 , 杨艺,等.基于体素双边注意力与八分体编码的点云实例分割[J].河南理工大学学报（自然科学版）,2026,45(5):87-96.

BI X H, NIU C B, YANG Y,et al.Point cloud instance segmentation based on voxel-based bilateral attention and octree feature encoding[J].Journal of Henan Polytechnic University(Natural Science) ,2026,45(5):87-96.

基于体素双边注意力与八分体编码的点云实例分割

毕晓华1,2,3, 牛春博2,3, 杨艺2,3, 王田4

1.焦作煤业集团有限公司，河南 焦作  454002；2.河南理工大学 电气工程与自动化学院，河南 焦作  454003；3.河南理工大学 河南省煤矿装备智能检测与控制重点实验室，河南 焦作  454003；4.北京航空航天大学 人工智能学院， 北京  100191

摘要:目的 如何在候选实例区域捕捉离群稀疏点云与目标主体之间的关联特征，生成精确的实例掩码，从而建立从位置空间到分类曲面的非线性映射模型，是点云实例精确分割的关键。  方法 本文构造了一种基于体素双边注意力机制的微型U网络，在体素层面融合点云通道和位置注意力，在关注位置局部特征的同时，通过学习各通道的重要性重新调整特征图，建立候选实例内部点云的远程依赖关系；同时，针对离群稀疏点云分类错误问题，提出一种八分体特征编码模块，将候选实例进行八分体拆分和编码，通过提取各分体的内部特性和分体间的关联特征，将远端离群散点纳入候选实例特征提取。结果 结果表明，本文提出的方法在mAP和mAP₅₀指标上的分类精度分别达到51.8%和68.2%，点云实例分割能力明显提升。结论 双边注意力机制和八分体特征编码模块能生成更加精确的实例掩码，有助于实现离群散点点云的准确分类。

关键词:点云实例分割;双边注意力;八分体特征编码;通道注意力

doi:10.16186/j.cnki.1673-9787.2025120086

基金项目:国家自然科学基金资助项目（92467108）

收稿日期:2025/12/24

修回日期:2026/03/27

出版日期:2026-07-24

Point cloud instance segmentation based on voxel-based bilateral attention and octree feature encoding

Bi Xiaohua1,2,3, Niu Chunbo2,3, Yang Yi2,3, Wang Tian4

1.Jiaozuo Coal Industry Group Co.， Ltd.， Jiaozuo  454002， Henan， China；2.School of Electrical Engineering and Automation， Henan Polytechnic University， Jiaozuo  454003， Henan， China；3.Henan Key Laboratory of Intelligent Inspection and Control of Coal Mine Equipment， Henan Polytechnic University， Jiaozuo  454003， Henan， China；4.School of Artificial Intelligence， Beihang University， Beijing  100191， China

Abstract: Objectives The essence of point cloud instance segmentation is to generate a closed classification surface in 3D space through a nonlinear mapping model， so that the target point cloud can be distinguished. However， the elongated and protruding point clouds of instance objects are usually distributed as outliers， which poses a challenge to accurate point cloud segmentation. Therefore， capturing the correlation features between outlier sparse point clouds and target objects in candidate instance regions， generating precise instance masks and thereby establishing a nonlinear mapping model from positional space to classification surfaces are considered critical for accurate point cloud instance segmentation.  Methods In this paper， a miniature U-Net based on the voxel-based bilateral attention mechanism was constructed. Channel and positional attention mechanisms of point clouds were fused at the voxel level， enabling the model to focus on local positional features while adaptively readjusting feature maps by learning the importance of each channel. In this way， long-range dependencies among point clouds within candidate instances were established. Meanwhile， to address the problem of outlier sparse point cloud misclassification， an octant feature encoding module was proposed. Candidate instances were split and encoded into octants， and both the internal characteristics of each octant and the correlation features between octants were extracted. As a result， distant outlier scattered points were incorporated into the feature extraction of candidate instances.  Results Experimental results show that the proposed method achieves 51.8% mAP and 68.2% mAP50， demonstrating a significant improvement in point cloud instance segmentation performance.  Conclusions More accurate instance masks were generated by the bilateral attention mechanism and octant feature encoding module， enabling precise classification of outlier point clouds.

Key words:point cloud instance segmentation;bilateral attention;octant feature encoding;channel attention