Reproduce “PIC-UIE: Predicting Image-Adaptive Corrections for Lightweight Underwater Image Enhancement” 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
Cunhao Zhu, Dongliang Xu, Xiangtao Kong, Xiaoyan Lu, Tianyu Wang, Yue Yao
PIC-UIE reframes underwater image enhancement as predicting a compact set of correction controls from a 256 × 256 thumbnail, then applying them to the original-resolution image. It reports 24.137 dB PSNR and 0.9216 SSIM on UIEB-90, with processing rates of 181.9 FPS at 1080p and 55.0 FPS at 4K on an RTX 4090; depth is used only as a training signal, not at inference.
Key insights
- Structuring the model to predict correction parameters rather than output pixels substantially outperformed similarly sized dense RGB and YCbCr residual predictors in the reported UIEB comparison.
- Using RGB for prediction and YCbCr for correction separates degradation cues from luminance and chroma adjustments; projected image-channel statistics and coupled chroma lookup tables also contribute to performance.
- The predicted transmission proxy correlated positively with relative depth in 88 of 90 UIEB-90 images, but this supports relative ordering rather than recovery of metric depth.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.