λ-domain VVC Rate Control Based on Game Theory

Jielian Lin, Aiping Huang, Keke Zhang, Xu Wang, Tiesong ZhaoarXiv 2022 · DOI · arXiv

In the evolving landscape of video streaming, the Versatile Video Coding (VVC) standard represents the state-of-the-art in compression efficiency. However, achieving optimal video quality under strict bandwidth constraints requires sophisticated rate control mechanisms. This paper proposes a novel approach to λ-domain rate control by integrating principles from Game Theory to manage bit allocation at the Coding Tree Unit (CTU) level.

The Challenge of Bit Allocation

Video compression is inherently a trade-off between bitrate and distortion. In VVC, λ-domain rate control typically uses mathematical models to relate these two variables. A persistent challenge is the distribution of a fixed bit budget across various regions of a single frame. Standard implementations often rely on heuristic weights or sliding windows, which may lack the theoretical rigor needed to ensure global quality optimization. If bits are allocated inefficiently—for instance, by over-allocating to static backgrounds at the expense of moving subjects—the overall visual experience suffers from inconsistent quality.

Methodology: CTUs as Cooperative Players

The core innovation of this research is treating each CTU within a frame as an independent agent or "player" in a Cooperative Bargaining Game.

Conceptual framework of CTU-level bit rate allocation using game theory and Nash Equilibrium.

The Nash Bargaining Solution (NBS)

To achieve a fair and optimal distribution, the authors define a utility function for each CTU as the inverse of its distortion ($1/D$). The bit allocation problem is then formulated to find the Nash Equilibrium. In this context, the goal is to reach a state where no single CTU can improve its quality without causing a disproportionate degradation in another, thereby maximizing the collective benefit of the entire frame.

Optimization Process

The algorithm employs a two-step optimization strategy:

  1. Newton Method Iteration: An iterative mathematical tool is used to solve for an intermediate variable ($\eta$) that represents the bargaining power or state of the game.
  2. λ Calculation: This variable is then used to derive the optimal λ value for each individual CTU.

Furthermore, the model accounts for "effective CTU sizes" to handle edge cases where CTUs might be smaller than the standard size, and it utilizes parameters from previously encoded frames to refine current estimates.

Experimental Results and Performance

The proposed method was rigorously evaluated using the JVET Common Test Conditions (CTC) across various resolutions (Class B through Class F) using the VTM 13.0 reference software in a Low-Delay B (LD) configuration.

Performance metrics (BDBR and BDPSNR) across different values of the scale factor ζ.

Key Findings

  • Bitrate Accuracy: The method achieved a significantly lower Rate Control Error (RCError) of 0.71%, compared to the default VTM's 1.36%, indicating much tighter adherence to target bitrates.
  • Compression Efficiency: In terms of BDBR (Bjøntegaard Delta Bit Rate), the game-theoretic approach outperformed the default VVC rate control by 0.48% and showed gains of up to 5.21% over other state-of-the-art algorithms.
  • Visual Quality: The average Y-PSNR reached 36.8579 dB, surpassing all benchmarked methods.
  • Low Overhead: Despite the complexity of game-theoretic modeling, the implementation is highly efficient, adding only 0.07% to the total encoding time.

Conclusion and Future Directions

By framing bit allocation as a cooperative game, this research provides a robust theoretical foundation for λ-domain rate control in VVC. The resulting Nash Equilibrium naturally balances quality across the frame, leading to superior rate-distortion performance. Future work may involve extending this model to Random Access (RA) configurations and integrating perceptual metrics like SSIM or VMAF into the utility function to better align with human visual perception.

Summary updated

BibTeX
@misc{lin2022domain,
  author = {Jielian Lin and Aiping Huang and Keke Zhang and Xu Wang and Tiesong Zhao},
  title = {{λ-domain VVC Rate Control Based on Game Theory}},
  year = {2022},
  eprint = {2205.03595},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2205.03595},
}
Figure