Polarization-based Camouflaged Object Detection

Abstract

Camouflaged Object Detection (COD) is a challenging task aiming to locate the “indistinguishable” object with high similarity to the background. Existing COD methods are highly dependent on the differences in extrinsic appearances while neglecting inherent physical characteristics. In this letter, we introduce polarization information to enlarge the discrepancies between objects and their surroundings for the COD task. We construct a polarization-based dataset PCOD that comprises 639 challenging real-world scenes, along with a CNN-based network PolarNet tailored to process polarization images. Our method achieves state-of-the-art performance on PCOD and other COD datasets, and shows enhanced qualitative and quantitative effectiveness compared with existing COD methods.