面向三维点云的端到端细粒度分类网络

An End-to-End Fine-Grained Classification Network for 3D Point Clouds

针对基于深度学习的三维点云分类方法在元类识别中对子类解译不足的问题,提出一种面向点云的三维模型细分类框架,构建具备层间语义相关和层内上下文感知的端到端细粒度点云网络——FGP-Net.首先以相互连接的卷积算子构建密集连接块,层内利用球邻域查询构造局部区域以完成局部到整体的特征映射,并通过偏置注意力机制关注上下文差异,从而更好地捕捉细粒度属性;然后在层间利用多层特征融合策略探索各层特征间的相关性,通过反向传播学习相关语义信息以提升模型分类性能.在FG3D的3个子数据集Airplane,Chair和Car上的实验结果表明,FGP-Net的总体准确率分别达到95.77%,80.88%和77.94%;与先进的三维点云分类模型PointNet++,PointCNN,Point2Sequence,DGCNN等相比,FGP-Net的分类性能均具有一定的优越性.

In terms of the problem of insufficient interpretation for sub-categories in meta-class recognition by 3D point clouds classification algorithms based on deep learning,a 3D model classification framework oriented to point clouds is proposed to build an end-to-end fine-grained point clouds network FGP-Net,which contains inter-layer semantic correlation and intra-layer context-awareness.That is,dense connections blocks are constructed by interconnected convolution operators,and local regions are constructed by the ball query in the layers to generate local regions for completing the feature mapping from local to global,then it passes through the offset-attention mechanism to pay attention to the contextual differences,so as to better capture fine-grained attributes.The multi-layer feature aggregation strategy is applied to explore the correlation between the features of each layer,and the relevant semantic information is learned through back-propagation to improve the classification performance of the model.On three sub-datasets of FG3D,Airplane,Chair and Car,the overall accuracy reaches 95.77%,80.88%and 77.94%respectively.Compared with the advanced 3D point clouds classification models PointNet++,PointCNN,Point2Sequence and DGCNN etc,FGP-Net has certain advantages in classification performance.