Polarimetric Underwater Stereo

Ryo Kawahara, Meng-Yu Jennifer Kuo, Takahiro Okabe

Abstract: We propose a novel method for recovering per-pixel surface normals and depth of objects in water. We show that these two geometric surface properties of each pixel can be reconstructed simultaneously when observing an object underwater, regardless of its texture. Our key idea is to leverage polarized rays along different underwater light paths while considering perspective projection and refraction. We implement our method using an off-the-shelf polarization camera and an LCD, and derive a practical calibration method to determine the values of its imaging parameters using a single planar mirror. Results on a number of real-world objects in water demonstrate the effectiveness of our method.