杨艺,牛春博,王科平,等.基于体素密度编码和类感知偏移的点云实例分割[J].河南理工大学学报（自然科学版）,2026,45(6):117-125.

Yang Y, Niu C B, Wang K P,et al.Point cloud instance segmentation based on voxel density coding and class aware offset[J].Journal of Henan Polytechnic University(Natural Science) ,2026,45(6):117-125.

基于体素密度编码和类感知偏移的点云实例分割

杨艺1,2, 牛春博1,2, 王科平1,2, 王田3

1.河南理工大学 电气工程与自动化学院，河南 焦作  454003；2.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；3.北京航空航天大学 人工智能研究院，北京  100191

摘要: 目的 为解决体素化容易丢失点云细节信息，且不同类别目标之间结构和尺度差异较大的问题，提出一种体素密度编码模块和一种类感知偏移机制，并与SoftGroup++结合构建性能更优的点云实例分割模型。  方法 首先，栅格下采样时，体素密度编码模块对所有体素栅格内原始点的密度信号进行并行编码，并与体素信息融合，补充体素细节信息、增强体素表征能力；其次，类感知偏移机制利用特定于类别的偏移系数，在偏移坐标空间中增强点分布的有序性，提升目标的实例中心聚类能力；最后，将上述2种机制集成到基于SoftGroup++改进的网络架构中，构建完整的点云实例分割模型。  结果 在大型室内数据集S3DIS和大型室外数据集STPLS3D上的实验结果表明，在数据集S3DIS上，本文方法的mAP和mAP₅₀分别达到58.8%，72.1%，对柱、窗户、门、沙发的分割精度在对比算法中达到了最优；在数据集STPLS3D上，本文方法的mAP和mAP₅₀分别达到51.4%，66.8%，在对比算法中性能最优。  结论 提出的体素密度编码模块和类感知偏移机制均可有效提升点云分割性能，且几乎不增加模型的推理时间，点云实例分割模型在室内室外场景下的分割精度均有明显提升。

关键词:三维点云;深度学习;实例分割;体素密度;偏移向量

doi:10.16186/j.cnki.1673-9787.2023120078

基金项目:国家自然科学基金资助项目（61972016）；国家重点研发计划项目（2018YFC0604502）；河南省科技攻关项目（232102210040）

收稿日期:2023/12/29

修回日期:2024/08/01

出版日期:2026/10/09

Point cloud instance segmentation based on voxel density coding and class aware offset

Yang Yi1,2, Niu Chunbo1,2, Wang Keping1,2, Wang Tian3

1.School of Electrical Engineering and Automation， Henan Polytechnic University ， Jiaozuo  454003， Henan，China；2.Henan Key Laboratory of Intelligent Detection and Control of Coal Mine Equipment ， Henan Polytechnic University， Jiaozuo  454003， Henan， China；3.Institute of Artificial Intelligence， Beihang University ， Beijing  100191， China

Abstract: Objectives To address the issues that voxelization tends to lose detailed point cloud information and that targets of different classes exhibit large variations in structure and scale， this paper proposes a voxel density coding module and a class-aware offset mechanism， which are integrated with SoftGroup++ to construct a point cloud instance segmentation model with improved performance.  Methods Specifically， during grid-based downsampling， the voxel density coding module encodes the density information of raw points within each voxel in parallel and fuses it with voxel features to enrich local details and enhance feature representation. In addition， the class-aware offset mechanism introduces category-specific offset coefficients to improve the spatial organization of points in the offset space， thereby enhancing instance center clustering. Finally， these two mechanisms are incorporated into an improved network architecture based on SoftGroup++ to build a complete point cloud instance segmentation model. Results Experimental results on the large-scale indoor dataset S3DIS and the large-scale outdoor dataset STPLS3D show that the proposed method achieves mAP/mAP50 of 58.8%/72.1% on S3DIS， with the best segmentation performance on categories such as columns， windows， doors， and sofas among the compared methods. On STPLS3D， the proposed method achieves mAP/mAP50 of 51.4%/66.8%， outperforming all compared methods.  Conclusions The proposed voxel density coding module and class-aware offset mechanism effectively improve point cloud segmentation performance while introducing negligible additional inference time. The resulting instance segmentation model achieves significant accuracy improvements in both indoor and outdoor scenarios.

Key words:3D point cloud; deep learning; instance segmentation; voxel density; offset vector