Point Cloud Geometry Coding with Relational Neighborhood Self-Attention

In the ever-evolving landscape of deep learning, attention models have contributed to boost the performance in diverse fields such as computer vision and natural language processing. Following this trend, this paper proposes a novel Relational Neighborhood Self-Attention (RNSA) model, specifically designed for Point Cloud (PC) geometry coding to be integrated in the emerging learning-based JPEG PCC standard. The RNSA model proposes three new methods: first, to effectively learn correlations between the points by capturing the relational features and positions of neighboring points; second, to address the inefficiencies of conventional dot product attention, a novel Relational Scoring method to generate an attention map able to capture both linear and non-linear relationships between points and their neighbors is adopted; third, the created attention maps are normalized by Sparsemax instead of Softmax to generate sparse probabilities and assigns higher scores to the most important neighbors while marginalizing the less significant ones. Experimental results show that the proposed attention model achieves around 8% gains in both BD-Rate PSNR Dl and PSNR D2 compared to the baseline codec, i.e., JPEG PCC, while adding a small number of model parameters to JPEG PCC.