基于视图融合和注意力机制的三维小目标识别

3D small objects recognition based on view fusion and attention mechanism

三维小目标识别是点云处理中的一个重要问题。针对小目标探测中点云稀疏导致网络提取特征少的问题,提出一种基于视图融合和注意力机制的点云识别网络。传统网络直接作用于单视角点云,其特征只包含目标的一部分,因此提出采用视图融合聚合不同视角下的局部点,以此增强全局特征。接着,针对聚合后的全局点特征提取困难的问题,提出采用注意力机制突出关键局部特征。实验结果表明,提出的方法在公开数据集ModelNet和真实环境中采集的Aircraft数据集上的效果都优于其它流行算法。

3D small objects recognition remains an important problem in point cloud processing.To solve the problem of sparse point clouds in small objects detection leading to few features extracted by the network,a point cloud recognition network based on view fusion and attention mechanism is proposed.Traditional networks directly act on a single view point cloud whose features contain only a portion of the target,so view fusion is proposed to enhance the global features by aggregating local points from different views.Experimental results show that the proposed method outperforms other popular algorithms on both the public dataset ModelNet and the Aircraft dataset collected in real environments.