DiffConv: Analyzing Irregular Point Clouds with an Irregular View

Standard spatial convolutions assume input data with a regular neighborhood structure. Existing methods typically generalize convolution to the irregular point cloud domain by fixing “view” through e.g. fixed size, where kernel size remains same for each point. However, since clouds are not as structured images, neighbor number gives an unfortunate inductive bias. We present novel graph named Difference Graph Convolution (diffConv), which does rely on view. diffConv operates spatially-varying and density-dilated neighborhoods, further adapted learned masked attention mechanism. Experiments show that our model is very robust noise, obtaining state-of-the-art performance in 3D shape classification scene understanding tasks, along faster inference speed. The code publicly available at https://github.com/mmmmimic/diffConvNet.