It is a challenge that design a deep neural network for raw point cloud, which is disordered and unstructured data. In this paper, we introduce a cross selfattention networks (CSANet) to solve raw point cloud classification and segmentation tasks. It has permutation invariance and can learn the coordinates and features of point cloud at the same time. To better capture features of different scales, a multi-scale fusion (MF) module is proposed, which can adaptively consider the information of different scales and establish a fast descent branch to bring richer gradient information to the network.
Note that during actual training, we set the model's learning rate to 0.01!