the sparse polarization data

[donyakh]

For the sparse polarization data, was a camera with sparse polarization sensors specifically manufactured for this study, or was the sparse data entirely generated through simulation? If the data was created by simulating sparse information from fully polarized data, how was the polarization effect on the RGB images reversed? my next question is regarding the noisy dataset. Why the noise is distributed just at the areas with a light source? why it's not randomly distributed?

[TeppeiKurita]

The data was generated through simulation. Specifically, the simulation includes data generated from real-world captured images as well as data created from computer graphics (CG). The CG data was produced using our custom polarization renderer.

Regarding the noisy dataset, the noise is simulated as shot noise based on the equation provided in the paper. Consequently, the noise is more pronounced in areas with a light source, as brighter regions tend to exhibit higher levels of noise.

[donyakh]

Thank you for replying to my comment. I have a question regarding the real captured data. In the supplementary material, for the raw image generation in the polarization camera dataset, only the output of a conventional polarization sensor is shown. However, to have ground truth data for machine learning, wouldn’t we need images taken with a camera equipped with a polarizer rotator?

In another work on sparse polarization, I came across this figure for sparse raw polarization data (Polarization image demosaicing and RGB image enhancement for a color polarization sparse focal plane array):

They were actually inspired by your work and used the same dataset from your study. I just wanted to confirm if, for the raw polarized sparse data, the images were captured using both the polarization sensor and an RGB camera equipped with a polarizer.