Despite deep convolutional networks have been highly successful in 2D vision tasks, extending them to point cloud classification task is still challenging due to irregularities of point cloud data relative to image grids. Recent methods usually take advantage of symmetric operators like max-pooling, to deal with point cloud order ambiguity. However, this kind of treatment does not consider the latent geometric information contained in the spatial order, which may limit the performance of feature learning. To address this issue, we present an adaptive permutation module (APM) that calculates a particular permutation from the input point clouds to achieve permutation invariance, as demonstrated by the visualization of the APM output feature maps. Thorough experiments are conducted to demonstrate the superiority of APM as well. In addition, the APM can be plugged into other state-of-the-art approaches flexibly to further improve performance in classification task. We build an end-to-end deep convolutional neural network applying PointCNN as our backbone combined with the adaptive permutation module and achieve state-of-the-art performance in point cloud classification task. Our work demonstrates that the latent spatially-local correlations play a critical role in feature learning on point clouds.

尽管深度卷积网络已在2D视觉任务中取得了巨大成功，但由于点云数据相对于图像网格的不规则性，将其扩展到点云分类任务仍然具有挑战性。最近的方法通常利用最大池等对称运算符来处理点云阶模糊性。但是，这种处理没有考虑空间顺序中包含的潜在几何信息，这可能会限制特征学习的性能。为了解决这个问题，我们提供了一个自适应排列模块（APM），该模块从输入点云计算特定排列以实现排列不变性，如APM输出特征图的可视化所示。进行了充分的实验以证明APM的优越性。此外，APM可以灵活地插入其他最新技术中，以进一步提高分类任务的性能。我们建立了一个以PointCNN为骨干与自适应置换模块相结合的端到端深度卷积神经网络，并在点云分类任务中实现了最新的性能。我们的工作表明，潜在的空间局部相关性在点云上的特征学习中起着至关重要的作用。