Polarized 3D: High-Quality Depth Sensing with Polarization CuesAchuta Kadambi, Vage Taamazyan, Boxin Shi, and Ramesh Raskar
Camera Culture Group, MIT Media Lab
Contribution
Take-home message:
Verification
Solution
Compare with other depth enhancement techniques
o Exploiting normals from polarization to enhance
the quality of a coarse depth map
o A physics-based framework, wherein the coarse
depth map is used to resolve azimuthal
ambiguity and correct for refractive distortion
o A spanning tree integration schemes,
specifically designed for polarization normals,
which uses the degree of polarization as a
weighting parameter
“Coarse depth map + normals from polarization” reconstructs 3D shape with various materials, under uncontrolled lighting condition, and with diffuse interreflection.
Low-cost depth sensing
Correcting normals from polarization
Depth + normal fusion
Shape from polarization
o A passive 3D camera capturing high-quality shape
o Robust to lighting, material variation, and diffuse interreflection
o Robust shape reconstruction fusing normals and depths
Full gradients Mini. Spanning Tree
Surface: GT Full integ. MST integ.
Depth: GT Kinect Pol. Ours Normal: GT Kinect Pol. Ours
o Robustness to various lighting condition
o Robustness to various material
o Robustness to diffuse interreflection
Resultso Uncontrolled lighting, complex object
Comparison with HQ Laser Scan