CANF-VC++: Enhancing Conditional Augmented Normalizing Flows for Video Compression with Advanced Techniques

Abstract

Video has become the predominant medium for information dissemination, driving the need for efficient video codecs. Recent advancements in learned video compression have shown promising results, surpassing traditional codecs in terms of coding efficiency. However, challenges remain in integrating fragmented techniques and incorporating new tools into existing codecs. In this paper, we comprehensively review the state-of-the-art CANF-VC codec and propose CANF-VC++, an enhanced version that addresses these challenges. We systematically explore architecture design, reference frame type, training procedure, and entropy coding efficiency, leading to substantial coding improvements. CANF-VC++ achieves significant Bj{\o}ntegaard-Delta rate savings on conventional datasets UVG, HEVC Class B and MCL-JCV, outperforming the baseline CANF-VC and even the H.266 reference software VTM. Our work demonstrates the potential of integrating advancements in video compression and serves as inspiration for future research in the field.