Point cloud geometry compression with sparse cascaded residuals and sparse attention

Effective compression of point clouds is essential for implementing virtual and mixed reality applications, which require encoding millions or even tens of millions of points. This paper offers a new geometric compression for point clouds based on sparse cascaded residuals and sparse attention. A sparse cascaded residual module is posited to connect multiple residual modules through shortcuts, thereby augmenting the network's learning capacity and compression efficacy. The authors developed a sparse attention module to acquire global features by computing interdependencies among points, enhancing compression performance to a greater extent. Trade-off parameters are employed to optimize the rate and distortion. The authors' method outperforms the state-of-the-art open-source method regarding rate-distortion on the ShapeNet, ModelNet, and Microsoft Voxelized Upper Bodies datasets, with average bjontegaard-delta (BD)-rate gains of -14.44% and -15.38%.