Reproduce “Generative Quanta Color Imaging” 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
Vishal Purohit, Junjie Luo, Yiheng Chi, Qi Guo, Stanley H. Chan, Qiang Qiu
This research demonstrates a method to generate a full-color image from a single 1-bit binary frame captured by a single-photon camera, addressing the challenge of extreme exposure variations inherent in such data. It achieves this by synthesizing a continuous spectrum of exposures via a Neural ODE with filter decomposition, followed by an exposure-adaptive colorization network, outperforming baselines in colorization quality and validated on real-world sensor data.
Key insights
- A single 1-bit binary frame from a single-photon sensor contains sufficient information to infer a full-color image, provided its inherent exposure variation is effectively managed.
- Neural ODEs, when combined with filter decomposition, offer a parameter-efficient and modular way to synthesize a continuous spectrum of image exposures from a single input, enabling robust pre-processing for generative tasks.
- Explicitly correcting for exposure variations before colorization, rather than relying on robust colorizers or extensive data augmentation, leads to superior and more consistent colorization performance for highly compressed imaging inputs.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.