基于自注意力机制的IW方法与3D-BoNet的实例分割网络

Instance Segmentation Network Based on IW Method of Self-attention Mechanism and 3D-BoNet Algorithm

针对实例分割算法中点云特征提取困难和鲁棒性低的问题,提出一种基于自注意力机制与3D-BoNet算法的实例分割网络(IW-BoNet)。在特征提取阶段,提出基于自注意力机制的Instance Wise(IW)方法,采用自注意力模块学习特征权重,捕捉实例上下文信息;将3D-BoNet模型中的欧式距离损失函数替换为Smooth L1损失函数。在STPLS3D数据集上的性能测试结果表明,与3D-BoNet模型相比,IW-BoNet模型平均均值精度提升6.2%,鲁棒性得到提升,能够更加高效地提取实例信息。

Aiming at the difficulty of point cloud feature extraction and low robustness in instance segmentation algorithms,an instance segmentation network(IW-BoNet)based on self-attention mechanism and 3D-BoNet algorithm was proposed.In the stage of feature extraction,a novel approach leveraging the selfattention mechanism,named of Instance Wise(IW),was proposed.The utilization of a self-attention module enabled effective learning of feature weights and facilitates capturing comprehensive contextual information pertaining to each instance.The Euclidean distance loss function in the 3D-BoNet model was replaced with the Smooth L1 loss function.The performance test on the STPLS3D dataset shows that compared with the original 3D-BoNet model,the average mean accuracy of IW-BoNet model is improved by 6.2%,and the robustness is improved,which can extract the instance information more efficiently.