Evaluating the effect of sparse convolutions on point cloud compression

Abstract

The use of point clouds as an imaging modality has been rapidly growing, motivating research on compression methods to enable efficient transmission and storage for many applications. While compression standards relying on conventional techniques such as planar projection and octree-based representation have been standardized by MPEG, recent research has demonstrated the potential of neural networks in achieving better rate-distortion performance for point cloud geometry coding. Early attempts in learning-based point cloud coding mostly relied on autoencoder architectures using dense convolutional layers, but the majority of recent research has shifted towards the use of sparse convolutions, which are applied only to occupied positions rather than the entire space. Since points are usually distributed on underlying surfaces rather than volumes, such operations allow to reduce the computational complexity required to compress and decompress point clouds. Moreover, recent solutions also achieve better compression efficiency, allocating fewer bits at similar levels of geometric distortion. However, it is not clear to which extent this gain in performance is due to the use of sparse convolutions, if any at all, since the architecture of the model is often modified. In this paper, we conduct an evaluation of the effect of replacing dense convolutions with sparse convolutions on the rate-distortion performance of the JPEG Pleno Point Cloud Verification Model. Results show that the use of sparse convolutions allows for an average BD-rate reduction of approximately 9% for both D1 and D2 PSNR metrics based on similar training procedures, with an even bigger reduction in point clouds featuring reduced point density.