NEURAL VIDEO COMPRESSION OVERVIEW, PERFORMANCE AND CHALLENGES
Marwa Tarchouli, Thomas Guionnet, Marc Rivière, Mickaël Raulet · 2025
In recent years, the emergence of neural compression has significantly disrupted the landscape of traditional codecs, surpassing contemporary standards within a short time. These codecs leverage two principal deep learning architectures. Variational Auto-Encoders (VAE) were firstly exploited due to their strong compatibility with compression tasks. Initially demonstrating exceptional performance in neural image coding, VAEs have been extended to encompass neural video coding. The second technology is Implicit Neural Representations (INR), aiming to reduce the complexity on the decoder side using an overfitted light model. This paper reviews the current state of neural video coding, examining both VAE and INR-based methodologies. It evaluates their effectiveness, highlights their strengths and limitations, and positions their performance against conventional codecs.