Efficient motion modelling with variable-sized blocks from hierarchical cuboidal partitioning

This paper explores the potential of cuboidal partitioning in motion modelling compared to the commonly used fixed-sized block-based architecture in scalable video coding. The traditional approach of dividing frames into fixed-sized blocks for independent motion compensation often results in coding inefficiency due to poor alignment with object boundaries. Hierarchical block partitioning has been introduced as a solution, but it suffers from an increased number of motion vectors, limiting its effectiveness. In contrast, cuboidal partitioning offers a promising alternative. It involves approximate segmentation of images into variable-sized rectangular segments (cuboids) that align well with object boundaries. The segmentation is based on a homogeneity constraint, minimizing the sum of squared errors (SSE). This property makes cuboidal partitioning compatible with block-based video coding techniques. In this paper, we investigate the potential of cuboids in motion modelling, specifically comparing them to fixed-sized blocks used in scalable video coding. Our approach involves constructing a motion-compensated current frame using the cuboidal partitioning information from the anchor frame within a group-of-pictures (GOP). The predicted current frame serves as the base layer, while the current frame is encoded as an enhancement layer using the scalable High Efficiency Video Coding (HEVC) encoder. Experimental results demonstrate significant bitrate savings ranging from 6.71% to 10.90% on 4 K video sequences. These savings highlight the superiority of our proposed model, which leverages cuboidal partitioning to improve coding efficiency and alignment with object boundaries. By adopting this approach, we mitigate the limitations of fixed-sized blocks and offer a more effective solution for motion modelling in scalable video coding.