Motion Vector Predictor Lists for Large-Motion Video Coding

The candidate list is segmented into multiple types of candidates (spatial candidates from neighboring blocks, temporal candidates from reference frames, history-based candidates from previously decoded blocks, object corresponding position candidates from semantic segmentation, and large motion candidates for high-motion regions). This segmentation allows the system to select appropriate candidate types based on scene characteristics, improving coding efficiency without uniformly increasing complexity across all blocks.

Object corresponding position candidates are derived in advance using semantic segmentation results before motion estimation. The semantic segmentation identifies objects and their positions in the current frame, and corresponding positions in reference frames are pre-calculated. This preliminary action provides high-quality motion vector predictors for objects with consistent semantics across frames, improving prediction accuracy especially for large motions.

The system dynamically adjusts the candidate list composition based on motion characteristics. For large motion regions, large motion candidates with scaled motion vectors are used; for stationary or small-motion regions, traditional spatial and temporal candidates suffice. This parameter change approach allows the encoder to adapt to different motion scenarios, maintaining high video quality while reducing the average data amount by avoiding unnecessary encoding of motion information for static regions.

Semantic segmentation results serve as an intermediary that guides the selection of object corresponding position candidates. The semantic segmentation identifies meaningful objects and their spatial-temporal relationships, enabling the system to select more accurate motion vector predictors for objects that move consistently across frames. This intermediary layer improves prediction accuracy without requiring full-frame motion estimation, reducing overall data requirements.

The candidate list composition is dynamic rather than static. The system adjusts which candidate types are included based on local motion characteristics, block position, and scene content. For example, object corresponding position candidates are only activated for blocks containing significant semantic objects, and large motion candidates are only used in regions with detected large motion. This dynamic adaptation improves prediction accuracy where needed while conserving computing resources in regions where traditional methods suffice.

Different candidate types are applied locally to different regions of the video frame based on their specific characteristics. Spatial candidates are used for neighboring blocks with similar motion, temporal candidates for blocks with consistent motion over time, object corresponding position candidates for semantically significant objects, and large motion candidates for high-motion regions. This local quality approach ensures high prediction accuracy for each region's specific motion pattern without uniformly increasing computing resources across the entire frame.