【摘要】 在三维视觉任务中,三维目标的未知旋转会给任务带来挑战,现有的部分神经网络框架对经过未知旋转后的三维目标进行识别或分割较为困难.针对上述问题,提出一种基于自监督学习方式的矢量型球面卷积网络,用于学习三维目标的旋转信息,以此来提升分类和分割任务的表现.首先,对三维点云信号进行球面采样,映射到单位球上;然后,使用矢量球面卷积网络提取旋转特征,同时将随机旋转后的三维点云信号输入相同结构的矢量球面卷积网络提取旋转特征,利用自监督网络训练学习旋转信息;最后,对随机旋转的三维目标进行目标分类实验和部分分割实验.实验表明,所设计的网络在测试数据随机旋转的情况下,在ModelNet40数据集上分类准确率提升75.75%,在ShapeNet数据集上分割效果显著,交并比(Intersection over union, IoU)提升51.48%.更多还原

【Abstract】 The unknown rotation of 3D objects can bring challenges to the 3D vision tasks. It is difficult for some existing neural networks to classify or segment the 3D model after the unknown rotation. Aiming at the above problems, this paper proposes a vector spherical convolutional network based on self-supervised for learning the rotation information of 3D objects. First, the 3D point cloud signal is spherically sampled and mapped to the unit sphere;then, the rotational features are extracted by the vector spherical convolution network, while the randomly rotated3D point cloud is input into the vector spherical convolution network of the same structure to extract rotation features, and the self-supervised network is used to train and learn the rotation information; finally, target classification experiments and part segmentation experiments are performed on randomly rotated 3D objects. Experimental results show that the network designed in this paper has a 75.75% improvement in classification accuracy on the ModelNet40 dataset and a significant segmentation effect on the ShapeNet dataset with a 51.48% improvement in the intersection over union(IoU) under random rotation of the test data.更多还原