Point Cloud Geometry Compression with Channel Attentions for Low Density Point Cloud

The shape of real-world objects is sensed and used in virtual space and other applications. Shapes are represented as a point cloud, and since the amount of data is enormous, deep learning based compression methods have been studied recently. Here, some point clouds have a high point density, while others have a low point density. However, coding schemes that take such point density into account have not been studied well. When the density of the point cloud is low, the spatial redundancy of the point cloud is low and the coding efficiency with recent convolution-based compression methods is also low. We propose a new method that combines convolutions with channel directional attentions to improve efficiency even when the point density is low.