Researchers have introduced PolarScale, a new benchmark designed to address the challenge of accurately estimating the radiometric scale in RGB-to-Stokes polarization imaging. Unlike previous methods that predict normalized Stokes components, PolarScale makes the scene's radiometric scale an explicit prediction target. This benchmark, built on existing full-Stokes measurements, evaluates semantic scale estimation against a constant-scale control, alongside other physical and angular consistency metrics. Experiments with various model architectures show that while explicit descriptor supervision improves accuracy, the learned scale's performance in tasks like diffuse/specular separation is comparable to a constant scale. AI
IMPACT Introduces a new benchmark for polarization imaging, potentially improving AI models' ability to reconstruct full Stokes representations from RGB inputs.
RANK_REASON The cluster describes a new benchmark for a computer vision task, presented in an arXiv paper. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
