Explorations on 3D point clouds coding using transformers and patches

As 3D point clouds become more common as a representation of 3D visual content, the need to efficiently compress these data grows ever stronger. Research has shown that deep learning based approaches to point cloud coding see an increase in performance when compared with competing encoders like those developed by MPEG. This article examines and evaluates the use of the Transformer architectures and patch-based inputs combined with well developed deep learning static point cloud compression solutions described in the literature. To that end, we propose four new deep learning encoders. The obtained results show an improvement over an octree based encoder proposed by MPEG and the baseline PCC Geo v2 codec in terms of a geometry fidelity metric. The article also presents an ablation study conducted to analyze the impact of several encoder related parameters and structures that can guide future research in deep learning point cloud compression.