Geometric Partition Motion Vector Refinement for Lower Signaling Overhead

The video block is divided into two geometric partitions with different motion characteristics. Each partition can independently apply motion vector refinement techniques (such as TMVD or MMVD) to improve motion accuracy locally without requiring full refinement across the entire block, thus balancing accuracy improvement with signaling overhead control.

Motion vector refinement is applied selectively rather than universally. The patent allows the decoder to choose whether to apply refinement techniques like TMVD or MMVD based on mode flags, enabling partial refinement only where needed to improve accuracy while avoiding unnecessary signaling overhead in regions where refinement is not beneficial.

The patent merges the advantages of geometric partition mode with existing merge modes (AMVP and merge). By integrating GPM into the merge framework and allowing selection through mode flags, the system achieves coding flexibility similar to non-merge modes while utilizing the efficient merge candidate structure to control signaling overhead.

The geometric partition mode is designed to work universally across different inter prediction modes (AMVP, merge, and skip). The same GPM structure and partitioning approach can be applied regardless of the base mode, providing coding flexibility through a single versatile mechanism rather than requiring separate specialized modes for each inter prediction type.