基于原型融合的小样本点云语义分割算法
Prototype fusion for few-shot point cloud semantic segmentation
冯站银 1方立 2艾国1
作者信息
- 1. 福州大学 先进制造学院,福建泉州 362200;中国科学院海西研究院泉州装备制造研究中心,福建泉州 362216
- 2. 中国科学院海西研究院泉州装备制造研究中心,福建泉州 362216
- 折叠
摘要
针对点云分割任务中标记样本不足的问题,提出一种基于原型融合的小样本点云语义分割算法.在骨干网络中,引入自适应权重用于捕获中心和相邻点之间的局部信息,选择性获取邻域特征;利用原型融合方法双向提取支持集和查询集的原型,对其进行融合、筛选获得融合后的原型用以分割.实验结果表明,与原型网络相比,该方法在数据集S3DIS和ScanNet上都表现出良好性能,验证了该方法具有良好的泛化能力.
Abstract
Aiming at the problem of insufficient labeled samples in the point cloud segmentation,a few-shot point cloud semantic segmentation algorithm based on prototype fusion was proposed.In the backbone network,an adaptive weight was designed to capture the local information between the center and neighboring points,selectively acquiring neighboring features.The proto-type fusion method was used to bidirectionally extract the prototypes of the support set and query set,those prototypes were fused and filtered to obtain the fused prototypes for segmentation.Experimental results show that,compared with the prototype network,the proposed method shows good performance on both S3 DIS and ScanNet datasets,verifying that this method has good generalization ability.
关键词
点云/小样本/点云分割/权重/邻域特征/原型融合/泛化能力Key words
point cloud/few-shot/point cloud semantic segmentation/weight/neighboring features/prototype fusion/generali-zation ability引用本文复制引用
基金项目
国家自然科学基金青年科学基金(42101359)
高层次人才创新创业基金(2020C003R)
出版年
2024
NETL
NSTL
万方数据