CNN-LNN Based Fast CU Partitioning Decision for VVC 3D Video Depth Map Intra Coding

Abstract

Currently, the coding efficacy of the cutting-edge video coding standard H.266/VVC surpasses that of 3D-HEVC (3D-High Efficiency Video Coding), but the existing VVC (Versatile Video Coding) low-complexity coding algorithm is mainly optimized for 2D video coding and cannot fully utilize the characteristics of the depth map itself. Based on this, we propose a fast decision algorithm employing the CNN (Convolutional Neural Network)-LNN (Lightweight Neural Network) model to diminish the intricacy of depth map intra coding in VVC 3D video. The algorithm treats the CU partitioning process in depth map coding as a two-stage process, first adding a non-local block and spatial pyramid pooling to the CNN model, enabling the proposed CNN model to skip the flat regions in the depth map and perform adaptive partitioning prediction of CUs in the edge regions; then, the LNN model is used to make early decision on TT (Ternary Tree) partition for CUs that need to be partitioned, and skip decisions for CUs that do not need to be partitioned by TT, so as to reduce some unnecessary RDO calculations. Experimental results illustrate that the algorithm achieves a notable reduction in encoding time amounting to 43.23% on average, with a negligible impact on the increase of BDBR.

Keywords

• VVC 3D video
• depth map coding
• CU early prediction
• CNN-LNN