Physics-guided Rolling Polarization Estimation (PRoPE)

Polarization imaging provides physical cues about material properties, reflections, and surface geometry that are unavailable from conventional intensity imaging. Rotating-polarizer systems preserve full spatial resolution and can acquire measurements over a dense range of analyzer orientations; however, reliable polarization reconstruction typically requires sufficient angular diversity, introducing acquisition latency that limits their applicability to dynamic scenes. We present Physics-guided Rolling Polarization Estimation (PRoPE), a framework that recovers the linear polarization state from only a few sequential measurements spanning a limited angular range. The key challenge is that Stokes estimation from closely spaced analyzer orientations is severely ill-conditioned and highly sensitive to measurement noise. PRoPE addresses this problem through a physics-guided decomposition that first refines the angle-independent intensity component, then analytically reconstructs the remaining polarization components using a better-conditioned formulation, followed by learned refinement of the recovered Stokes representation. This design enables polarization estimates to be continuously updated from a short sliding window of measurements, avoiding the need to complete a wide angular sweep before reconstruction. We evaluate PRoPE on 529 indoor and outdoor scenes acquired over a dense 180° analyzer sweep. On the 52-scene test set, PRoPE achieves a DoLP PSNR of 29.336 dB and an AoLP mean angular error of 15.317° using only three sequential measurements at 15° angular intervals. These results demonstrate that accurate polarization reconstruction is possible from a substantially reduced angular acquisition range, supporting lower-latency polarization sensing for dynamic vision and robotic perception.