Context-Based Coding for Non-Affine AMVR Motion Vector Resolution

The patent applies adaptive motion vector resolution (AMVR) that dynamically changes the precision parameter of motion vectors based on block characteristics. For blocks with small motion variations, lower precision (e.g., integer pixel) is used, while blocks with large motion variations use higher precision (e.g., quarter-pixel). This parameter adaptation resolves the contradiction by maintaining high accuracy where needed while reducing computational complexity where high precision is unnecessary.

The patent implements different motion vector precision levels for different regions within the video frame based on local motion characteristics. Affine motion models are applied selectively to blocks exhibiting affine motion patterns, while other blocks use simpler prediction methods. This local differentiation maintains high motion prediction accuracy for complex regions while reducing overall computational complexity.

The patent segments the context modeling process into multiple independent context models based on the AMVR mode type. Different context models are maintained for affine AMVR blocks versus non-affine AMVR blocks, allowing each context model to be optimized independently. This segmentation improves coding efficiency by accurately predicting mode probabilities while managing context model complexity through modular organization.

The patent uses neighboring block AMVR mode information selectively rather than universally. Context modeling leverages neighboring block information when it provides meaningful prediction, but avoids using it when neighboring blocks are unavailable or when the additional complexity does not provide sufficient coding gain. This partial application resolves the contradiction between coding efficiency and model complexity.