### Plenoptic Sampling and Rendering

```Plenoptic Sampling,
Rendering and
Super-Demosaicing
Todor Georgiev, Qualcomm
Sergio Goma, Qualcomm
Zhan Yu, University of Delaware
Andrew Lumsdaine, Indiana University
Overview
o Sampling: The Lippmann Sensor
1.0 vs. 2.0 sampling
o Rendering, focusing, polarization
o Super-Resolution
1. The light field (radiance) is a function in 4D ray space, not a field in 3D:
2. The first interesting problem: Continuity
A 4D function can’t be represented continuously by a 2D function.
Radiance capture must be discontinuous. Multiplexing with microlenses.
Small (10 pixels) microimages have significant losses at the edges.
Sampling the radiance: 1.0 vs. 2.0
Lippmann proposed sampling with an array of microlenses (the Lippmann sensor):
“Integral Photography”
Two distinctly different versions. Will compare resolution and show the key difference.
Plenoptic 1.0 sensor is sampling radiance at the microlenses:
Plenoptic 2.0 sensor is sampling at a distance a in front of the microlenses:
Δp
Comparing resolution
Plenoptic 1.0: sampling at the pitch of the microlenses, d.
Comparing resolution
Plenoptic 1.0: sampling at the pitch of the microlenses, d.
Plenoptic 2.0: sampling at δa/b, where δ is the pixel size.
Comparing resolution
Plenoptic 1.0: sampling at the pitch of the microlenses, d.
Plenoptic 2.0: sampling at δa/b, where δ is the pixel size.
When is 2.0 better than 1.0?
Comparing resolution
Plenoptic 1.0: sampling at the pitch of the microlenses, d.
Plenoptic 2.0: sampling at δa/b, where δ is the pixel size.
When is 2.0 better than 1.0?
This is equivalent to
- This formula shows the distinction 1.0/2.0
- Resolution improves as you increase b.
Rendering is focusing
Images are rendered by “integral projection”, i.e. adding up all pixels along a given
direction. This is same as focusing.
Rendering is focusing
Images are rendered by “integral projection”, i.e. adding up all pixels along a given
direction. This is same as focusing. Changing the direction of projection changes
the plane of focus. (Ren Ng, 2005)
Rendering is focusing
Images are rendered by “integral projection”, i.e. adding up all pixels along a given
direction. This is same as focusing. Changing the direction of projection changes
the plane of focus. (Ren Ng, 2005)
This process is not really re-focusing, but rendering by projection: focusing.
HDR and Polarization sampling/rendering
Extended modalities like HDR, polarization and others can be sampled with
filters on the microlenses, or on the main camera lens. One example:
Difference in functionality between 1.0 and 2.0.
Demo: light reflected off water is partially polarized…
Kodacolor and Polarization
Kodacolor
Plenoptic 1.0 camera capturing color was made by Kodak (1928)
Using black and white film
1.0 Super-Resolution
In both 1.0 and 2.0 we superresolve the same way:
- integral projection
This interleaves pixel from neighboring microimages, increasing resolution:
In 1.0 we need to shear a lot more than in 2.0, also need to deconvolve.
2.0 Super-Resolution
In 2.0 camera we shear less. That’s because radiance is captured sheared.
Example with no shear:
The difference: In 2.0 pixels are already in focus. They “fit in” exactly.
No need for deconvolution. Because of that quality is better.
Super-Demosaic
The color filters from the Bayer array are equivalent to spacing between pixels
in each individual color channel.
Algorithm:
o First effectively shear the raw data to the right location
o Then do integral projection and demosaic
This gains you 4 times better resolution compared to