Reproduce “IFRNet: Intermediate Feature Refine Network for Efficient Frame Interpolation” 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
Lingtong Kong, Boyuan Jiang, Donghao Luo, Wenqing Chu, Xiaoming Huang, Ying Tai, Chengjie Wang, Jie Yang
IFRNet introduces an efficient network for video frame interpolation that jointly refines intermediate flow fields and features within a single encoder-decoder, achieving state-of-the-art accuracy with significantly faster inference speeds and fewer parameters than previous methods. It improves PSNR on Vimeo90K by 0.63 dB compared to CDFI full while being 15.2 times faster, making high-quality VFI practical for real-time applications.
Key insights
- Jointly refining intermediate flow and features within a single network enhances both motion estimation and frame synthesis, leading to better quality and efficiency than separate, cascaded modules.
- Task-oriented and adaptively robust flow distillation effectively transfers useful knowledge from teacher models while selectively ignoring noise, crucial for training VFI networks with imperfect pseudo-labels.
- Regularizing intermediate features in a multi-scale census space (Feature Space Geometry Consistency Loss) helps preserve structural details and layout, which is vital for synthesizing high-quality, sharp frames.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.