联合边界感知和多特征融合的点云语义分割方法

Joint boundary-aware and multi-feature fusion for point cloud semantic segmentation

针对当前大多数基于深度学习的点云语义分割方法容易忽略过渡区域的目标边界,造成边界处存在模糊特征的问题,提出了一种联合边界感知和多特征融合(boundary-aware and multi-feature fusion, BA-MFF)的点云语义分割方法。首先,对骨干网络进行优化,使得提取到的特征更具有鲁棒性;其次,设计了边界感知模块(boundary-aware module, BAM)关注过渡区域的目标边界,该模块包含边界点预测模块(boundary point prediction module, BPPM)和特征聚合模块(feature aggregation module, FAM);边界点预测模块通过学习邻域点特征预测出属于边界上的点,特征聚合模块在邻域内对点云特征进行判别聚合;最后,为获得更有鉴别性的特征,引入了多特征融合模块(multi-feature fusion module, MFFM)对不同通道之间的特征进行了融合。实验结果表明:该方法在ScanNetV2数据集上平均交并比(mean intersection over union, mIoU)达到63.7%,在S3DIS数据集上总体精度(overall accuracy, OA)和mIoU分别为88.2%和62.3%。该方法有效关注了过渡区域,具有一定的分割优越性。

Current point cloud semantic segmentation methods based on deep learning tend to o-verlook the boundary of objects in transition area,resulting in the problem of ambiguous features at the boundary.This article proposed a point cloud semantic segmentation method with bounda-ry-aware and multi-feature fusion(BA-MFF).Firstly,the backbone network was optimized to make the extracted features more robust.Secondly,boundary-aware module(BAM)was de-signed to focus on the boundaries of objects in transition area.This module consisted of a bound-ary point prediction module(BPPM)and a feature aggregation module(FAM).The boundary point prediction module predicted the points belonging to boundaries by learning the features of neighboring points,and the feature aggregation module performed discriminative aggregation of point cloud features within the neighborhood.Finally,for more discriminative features,a multi-feature fusion module(MFFM)was introduced,which fused features between different chan-nels.The experimental results show that the mean Intersection over Union(mIoU)of this meth-od reaches 63.7%on the ScanNetV2 dataset.The Overall Accuracy(OA)and mIoU on the S3DIS are 88.2% and 62.3%,respectively.The method in this paper effectively focuses on the transition area and has some segmentation superiority.