Reproduce “Frequency-aware Neural Representation for Videos” 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

Jun Zhu, Xinfeng Zhang, Lv Tang, Junhao Jiang, Gai Zhang, Jia Wang

Researchers at the University of Chinese Academy of Sciences developed FaNeRV, a frequency-aware neural representation for videos that directly addresses the spectral bias problem in implicit neural representations. This framework achieves superior rate-distortion performance, surpassing the H.266/VTM standard with a 6.1% PSNR BD-Rate saving on the HEVC ClassB dataset, while also offering scalable coding capabilities.

Key insights

  • Explicitly decoupling low and high-frequency learning through multi-resolution supervision and dynamic high-frequency injection effectively mitigates spectral bias in INRs.
  • Allocating more network capacity to shallow layers for low-frequency modeling provides a stable structural prior, enabling deeper layers to focus more effectively on high-frequency details.
  • A hybrid upsampling approach optimally balances efficient recovery of global structures with targeted high-frequency residual capture, enhancing both reconstruction quality and coding efficiency.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.