LRC-DHVC: Towards Local Rate Control in Neural Video Compression

TL;DR

Problem: Local rate control in neural video compression is underdeveloped, limiting how quality is distributed spatially within frames. Approach: LRC-DHVC extends the DHVC framework by conditioning on a pixel-level quality map $m_t \in [0,1]$ and introducing a per-latent weight $\boldsymbol{\omega}$ to scale latent representations before quantization and after decoding; the quality map is mapped to distortion weights via $\boldsymbol{\Lambda}_t = \alpha \cdot e^{\beta \cdot m_t}$ with $\alpha=0.001$, $\beta=6$, and the loss is ${\cal L} = {\cal L}_{rate} + {\cal L}_{W-MSE}$ with ${\cal L}_{W-MSE} = \frac{1}{H W} \sum \boldsymbol{\Lambda}_t \odot (x_t - \hat{x}_t)^2$. Findings: a single model spans wide rate-distortion ranges with constant parameter count, and optimized quality maps substantially close the BD-rate gap to single-rate baselines on unseen data, validating pixel-level rate control for machines. Significance: enables practical region-specific RD control for machine-oriented video coding and can be extended to other neural compression architectures.

Abstract

Local rate control is a key enabler to generalize image and video compression for dedicated challenges, such as video coding for machines. While traditional hybrid video coding can easily adapt the local rate-distortion trade-off by changing the local quantization parameter, no such approach is currently available for learning-based video compression. In this paper, we propose LRC-DHVC, a hierarchical video compression network, which allows continuous local rate control on a pixel level to vary the spatial quality distribution within individual video frames. This is achieved by concatenating a quality map to the input frame and applying a weighted MSE loss which matches the pixelwise trade-off factors in the quality map. During training, the model sees a variety of quality maps due to a constrained-random generation. Our model is the first neural video compression network, which can continuously and spatially adapt to varying quality constraints. Due to the wide quality and bit rate range, a single set of network parameters is sufficient. Compared to single rate point networks, which scale linearly with the number of rate points, the memory requirements for our network parameters remain constant. The code and model are available at link-updated-upon-acceptance.

Figures (5)

• Figure 1: Architectural overview of LRC-DHVC, our hierarchical video codec for local rate control using a pixel-granular quality map $\bm{m}_t$. The latent block incorporates temporal modeling via the temporal priors $\bm{F}^l_{<t}$, probability modeling, quantization and entropy coding. Novel components are highlighted in red.
• Figure 2: Architectural overview of the latent block, where modified components are highlighted in red.
• Figure 3: Exemplary compressed frame with corresponding quality map (white: high quality) and bit distribution heatmap. Best to be viewed enlarged on a screen.
• Figure 4: Rate-distortion curves comparing single-rate models for DHVC lu2024 with our LRC-DHVC which integrates local rate control in a single model. The continuous RD-curve of LRC-DHVC (Uniform) is approximated by the evaluation of 21 rate points ($\bm{m}_t=\{0, 0.05, ..., 0.95, 1.0\}$). The eight distinct quality maps for LRC-DHVC (Optimized) are derived from the hyperparameter set ($\lambda=\{64, 128, 256, 512, 1024, 2048, 4096, 8192\}$).
• Figure 5: Relation between number of rate points and number of parameters for DHVC lu2024 and LRC-DHVC. The parameter count for LRC-DHVC remains constant, whereas it increases linearly for DHVC.