Dual-Level Adaptive RDO for Learned Wavelet Video Coding

Abstract

End-to-end learned video compression has recently demonstrated the potential to surpass traditional standards like HEVC and VVC. Among existing paradigms, learned wavelet coding using motion-compensated temporal filtering (MCTF) is particularly attractive for its inherent scalability and interpretable latent representations. However, current MCTF frameworks typically rely on a uniform Lagrange multiplier λ schedule, ignoring the heterogeneous importance of temporal layers and frame-pair content variability, which leads to suboptimal bit allocation. We propose pMCTF-DRD, a structure-and content-aware dual-level RDO framework for learned wavelet video coding. Our approach decouples RDO into two granularities: a Cross-Level λ Allocator that explicitly models temporal layer importance to enforce hierarchical quality control during inference, and a fine-grained content-aware RD controller that predicts adaptive ∆λ adjustments for high-pass frame pairs based on localized motion, residual, and texture features. These modules are plug-and-play and can be seamlessly integrated into existing MCTF-based codecs without any backbone retraining. Experimental results demonstrate that pMCTF-DRD achieves over 10% BD-rate reduction compared to state-of-the-art baselines while significantly improving reconstruction fidelity across diverse video content.