Affine Merge Candidate Selection for Video Compression

The patent segments the predictor selection process into multiple independent stages: first identifying a restricted set of causal neighboring CU candidates, then evaluating rate distortion metrics for each candidate, and finally selecting the optimal predictor. This segmentation allows the system to maintain low complexity by limiting the search space while still achieving good rate distortion performance through systematic evaluation of the segmented candidate set.

The patent performs preliminary action by pre-identifying and restricting the set of candidate predictors to only causal neighboring CUs before the actual predictor selection process. This preliminary restriction of the candidate pool reduces the computational complexity of the subsequent rate distortion evaluation, while still maintaining sufficient candidate diversity to achieve good rate distortion performance.

The patent applies local quality by considering different types of candidate CUs (spatial neighbors, temporal neighbors, co-located blocks) with different weights and priorities in the predictor selection process. Rather than uniformly evaluating all possible candidates, the system focuses computational resources on locally relevant candidates that are more likely to provide accurate motion prediction, thus improving rate distortion performance without proportionally increasing complexity.

The patent introduces dynamics by adaptively adjusting the set of candidate predictors based on the specific video content and motion characteristics being encoded. The candidate pool is dynamically constructed from available causal neighboring CUs, allowing the system to expand or contract the search space depending on the local motion patterns, thereby achieving good rate distortion performance with variable complexity rather than fixed high complexity.