ASIS模块支持下融合注意力机制KNN的点云实例分割算法

Point cloud instance segmentation based on attention mechanism KNN and ASIS module

针对基于3D卷积的点云实例分割算法的分割结果离散化、特征利用不充分的问题,提出具有注意力机制(KNN)模块和改进的实例语义关联(ASIS)模块的点云实例分割模型.模型以体素作为输入,通过3D子流形稀疏卷积提取点特征.利用具有注意力机制的KNN算法,对语义、实例特征空间的特征进行重组,以缓解提取到的特征离散化问题.通过改进的ASIS模块,对重组后的语义、实例特征相互关联以增强点特征间的区分度.对于语义特征与实例嵌入,分别应用Softmax模块、MeanShift算法获得语义与实例分割结果,采用S3DIS公开数据集对所提模型进行验证.实验结果表明,所提模型的实例分割结果在平均实例覆盖率(mCov)、平均加权实例覆盖率(mWCov)、平均精确率(mPrec)、平均召回率(mRec)衡量指标上分别达到了53.1%、57.1%、65.2%与52.8%;语义分割结果在平均交并比和总体精度上分别达到了61.7%、88.1%.消融实验结果验证了所提模块的有效性.

A point cloud instance segmentation model with a k-nearest neighbors(KNN)module featuring attention mechanism and an improved associatively segmenting instances and semantics(ASIS)module was proposed to address the problems of discrete segmentation and insufficient feature utilization in traditional 3D convolution-based algorithms.The model took voxels as input and extracted point features through sparse convolution of 3D submanifolds.The KNN algorithm with attention mechanism was used for reorganizing the features in the semantic and instance feature space to alleviate the problem caused by the quantization error of extracted features.The reorganized semantic and instance features were correlated through the improved ASIS module to enhance the discrimination between point features.For semantic features and instance embedding,the softmax module and the meanshift algorithm were applied to obtain semantic and instance segmentation results respectively.The public S3DIS dataset was employed to validate the proposed model.The experimental results showed that the instance segmentation results of the proposed model achieved 53.1%,57.1%,65.2%and 52.8%in terms of mean coverage(mCoV),mean weighted coverage(mWCov),mean precision(mPrec)and mean recall(mRec)for the instance segmentation.The semantic segmentation achieved 61.7%and 88.1%respectively in terms of mean intersection over union(mIoU)and Over-all accuracy(Oacc)for the semantic segmentation.The ablation experiment verified the effectiveness of the proposed modules.