Reproduce “Rethinking Generative Human Video Coding with Implicit Motion Transformation” 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
Bolin Chen, Ru-Ling Liao, Jie Chen, Yan Ye
Researchers at Alibaba DAMO Academy developed Implicit Motion Transformation (IMT), a new paradigm for generative video coding that implicitly models motion dynamics within compact feature representations for human body videos. IMT achieves up to 72.46% bitrate savings over VVC while producing perceptually superior reconstructed video sequences.
Key insights
- Motion information can be effectively encoded as implicit transformations within learned feature representations, capturing higher-level motion semantics.
- Implicit motion modeling, leveraging cross-attention, addresses limitations of explicit motion estimation in complex human body videos.
- Direct GAN-based generation from transformed features avoids motion estimation errors common in warp-based reconstruction schemes.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.