Inter predictive coding for point cloud attributes with online coordinate alignment and multi-scale latent prediction

The existing deep learning (DL) based compression for the attribute information of point cloud is normally designed to remove the spatial redundancy within a single frame. It cannot remove the temporal redundancy between point cloud frames, thus resulting in limited compression efficiency. To solve this problem, we introduce inter predictive coding in the DL-based point cloud attribute compression to improve the coding efficiency. This predictive coding occurs in the latent space, where we generate the multi-scale predicted latent representations for the attribute information, with the guidance of coordinate information. The final predicted representation is obtained by combining all the predicted ones with different scales. Before the generation of predicted representations, we align the coordinates of the reference and target point cloud frames in an online manner, to effectively eliminate the misalignment between them for the accurate prediction. Experimental results demonstrate that our proposed method outperforms the state-of-the-art DL-based attribute compression method, namely TSC-PCAC, achieving around 24.10% BD-BR reduction (on average) and 0.77 dB average BD-PSNR gain (on average).