Abstract
The Versatile Video Coding (VVC) standard achieves significant compression gains over its predecessor, HEVC, but introduces a massive increase in encoding complexity. This paper presents a detailed statistical analysis of the VVC Inter coding process using the VTM-15.0 test model. By identifying "low-yield" coding units and modes—those that consume high computational resources but are rarely selected—the study highlights specific targets for future encoder optimization and AI-driven acceleration.
Introduction
The 2020 VVC standard offers a 50% improvement in compression efficiency over HEVC. However, this comes at the cost of encoding times that are 5 to 37 times longer. A primary driver of this complexity is the Rate-Distortion Optimization (RDO) process, which exhaustively evaluates numerous block partitions and prediction modes. This research aims to profile the RDO process to find computational bottlenecks where the search effort does not justify the marginal gains in compression.
Methodology
The study utilized the VVC Test Model (VTM-15.0) and deactivated built-in early-exit heuristics to measure the full search space of the RDO process. The researchers analyzed 10 Inter coding modes and a variety of Coding Unit (CU) sizes across standard JVET Common Test Condition (CTC) sequences (720p to 4K).
Key Insights and Results
The analysis revealed that the exhaustive search space is often inefficient. Deactivating shortcuts increased complexity by 138% while only improving compression efficiency (BD-rate) by 0.76%.
CU Size Inefficiency
Certain CU sizes, particularly 16x8, 8x16, and 16x16, were identified as "low-yield." For example, 16x16 blocks often require as much processing as 32x32 blocks but are selected as the final winner only half as often. Conversely, larger blocks (64x64 and 128x128) account for roughly one-third of total complexity, with their selection frequency increasing significantly at higher Quantization Parameters (QP).
Mode Selection and QP Sensitivity
The selection of coding modes is highly dependent on the QP setting. At lower quality (higher QP), the encoder favors simpler modes like Skip.
Advanced VVC tools like AffineMerge and AffineSkip are significant time-consumers. Despite their complexity, AffineMerge is selected only 1.9% of the time for 128x128 blocks, suggesting that skipping these tests for large blocks could provide immediate speedups with negligible quality loss.
Conclusion
This statistical profiling identifies clear "red flags" in the VVC RDO process. Future work will focus on designing adaptive algorithms that use these probabilities to skip unlikely modes or partitions, paving the way for more efficient, real-time VVC implementations.
Summary updated