New fast MMVD algorithm cuts VVenC encoding complexity by half

Merge with Motion Vector Difference (MMVD) is a key tool in Versatile Video Coding (VVC) that improves motion prediction accuracy, but its exhaustive search heavily burdens encoders. A team of researchers (Feng, Xu, Chen, Li, Liu, Wu) from academia addresses this by introducing a fast MMVD algorithm for VVenC, the open-source VVC encoder. Their method approximates the computationally expensive 8-tap interpolation filter used in fractional motion estimation with a simpler 2-tap filter. This approximation enables derivation of a criterion based on spatial gradients and prediction residuals to estimate the potential gain of MMVD candidates, pruning unpromising ones without full search. The team generalizes the criterion for both shifted integer reference samples and 2D separable filtering.

To minimize implementation overhead, optimizations include symmetric offset inference (exploiting motion vector symmetry) and cross-shaped downsampled dot-product computation. Compared to existing fast MMVD algorithms in VVenC, the proposed method reduces average MMVD search ratio from 21.07% to 11.05% — a near 48% reduction in candidate evaluations. The efficiency-complexity metric η, which balances coding efficiency gains against runtime cost, improved from 11.79 to 7.10 under the fast preset. The work is documented in a 5-page paper with 4 tables and 4 figures, currently on arXiv (eess.IV/2606.31084). Though not yet integrated into the main VVenC branch, the algorithm promises significant speedups for real-time and power-constrained encoding scenarios.