Reducing the Complexity of Normalizing Flow Architectures for Point Cloud Attribute Compression

"Learning-based point cloud (PC) compression is a promising research avenue to reduce the transmission and storage costs for PC applications. Existing learning-based methods to compress PCs attributes employ variational autoencoders (VAE) or normalizing flows (NF) to learn compact signal representations. However, VAEs leverage a lower-dimensional bottleneck that limit the maximum reconstruction quality, while existing NF architectures have a big complexity in terms of number of coefficients in the network. In this paper, we propose an improved NF architecture combining two models to compress PC attributes. Our combined approach reduce the number of parameters of the existing NF architectures by over 6x and achieves state-of-the-art coding gains compared to previous learning-based methods, with, in some cases, a comparable performance to G-PCC v.21, showing the potential of this scheme for PC compression."