Demo: RollingPol: Single-Shot Polarization Imaging for Underwater Sensing
Abstract
Underwater optical sensing is severely degraded by backscattered light from suspended particles. While polarization-based image enhancement methods can suppress this backscattering effect, existing methods are either costly, or require sequential image captures that can't handle dynamic scenes. We address this challenge with RollingPol, a single-shot polarization imaging system that couples fast polarization modulation with the rolling-shutter readout of commodity cameras to encode multiple polarization states within a single exposure. A neural network is then used to reconstruct images at given polarization states from one RollingPol capture, enabling conventional polarization-based enhancement without sequential acquisition. The resulting physics-grounded enhancement can be used directly or further supplied as a prior to existing learning-based underwater image enhancement models. We demonstrate that our pipeline not only improves underwater visibility and suppresses backscatter, but also enhances the performance of downstream learning-based methods across diverse conditions.