Reproduce “Region-aware Adaptive Instance Normalization for Image Harmonization” 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 Ling, Han Xue, Li Song, Rong Xie, Xiao Gu
Researchers from Shanghai Jiao Tong University introduce the Region-aware Adaptive Instance Normalization (RAIN) module, reframing image harmonization as a background-to-foreground style transfer challenge. This module explicitly extracts visual style parameters from the background to adaptively adjust foreground features, leading to an average PSNR of 36.12 dB on the iHarmony4 dataset, a 2.2 dB improvement over a baseline model.
Key insights
- Explicitly framing image harmonization as a background-to-foreground style transfer problem, rather than relying on implicit domain alignment, yields superior visual consistency.
- The statistical properties of background features can effectively serve as adaptive 'style parameters' to guide the transformation of foreground features.
- Placing RAIN modules in the network's decoder, where image synthesis occurs, is more effective for style transfer, suggesting that explicit style adaptation is crucial during the feature reconstruction phase.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.