Reproduce “Cool-chic video: Learned video coding with 800 parameters” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Thomas Leguay, Théo Ladune, Pierrick Philippe, Olivier Déforges
Orange Innovation and IETR researchers developed Cool-chic video, an overfitted learned video codec that extends the Cool-chic image codec with an inter-coding module. This system achieves exceptionally low decoding complexity, reporting 807 parameters and 924 multiplications per decoded pixel, while offering competitive rate-distortion performance against H.264/AVC at low bitrates.
Key insights
- Overfitted latent feature maps can serve as an efficient primary compressed representation for video, maintaining spatial organization suitable for temporal redundancy exploitation, unlike approaches compressing network parameters directly.
- A lightweight inter-coding module, devoid of learned parameters and driven by optimized signals from the learned decoder, can effectively manage temporal redundancies, including motion compensation and residual coding.
- Pixel-wise prediction mode and bidirectional prediction weighting are crucial for optimizing rate-distortion, allowing the system to selectively apply temporal prediction or blend reference frames as needed.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.