Reports

Contents

Title: Picam FWB Rescaling - MB
Date:2022-02-18 - 2022-11-03
Data File: MB_PicamResc_AWB.csv
MB_PicamResc_R14B13.csv
MB_PicamResc_R18B18.csv
Refers to:MB, RPI0W-003

Acronyms

R = red

B = blue

G = green

GCC = green chromatic coordinate, [green/(red+blue+green)]

AWB = automatic white balance where red, blue, and green gains are set by the camera

FWB = fixed white balance where at least one of the red, blue, or green gains is set to a fixed value by the user

 

Mayberry

2021-04-27 (DOY 116): Installed picam onto MB tower for initial camera comparison with Canon and Stardot camera

2021-09-15 (DOY 258): Added FWB photos using R=1.4, B=1.3 (scaling factors Joe found online)

2022-06-08 (DOY 159): Changed FWB to use R=1.8,B=1.8 (scaling factors determined by Joe's analysis comparing R1.4B1.3 photos with AWB photos).

 

The FWB photos are nicer because there is less noise in the GCC values day to day. However, with our current RGB gains for FWB photos, there is a step change in the timeseries graph of GCC values when the data changes from AWB to FWB. To make a nice-looking graph, we decided to scale the FWB photos so the GCC from FWB photos is about the same magnitude as the GCC from the AWB photos.

To calculate rescaling factors, I used linear regression comparing R, G, and B between AWB and FWB photos for all daytime data. Restricting the regression to midday data did not provide enough range for a good regression. I forced the regression through 0 and used the slope of the regression line as my rescaling factor.

Joe did some initial analysis with the FWB photos when R=1.4, B=1.3 and determined that if we scaled R and B so that R=1.8, B=1.8, the resulting GCC would be the same magnitude as the AWB GCC.

I also calculated rescaling factors for R, G, B using linear regression comparing R, G, B between AWB and FWB photos for all daytime data. Restricting the regression to midday data did not provide enough range for a good regression. I forced the regression through 0 and used the slope of the regression line as my rescaling factor.

 

DOY 2021-258 to 2022-159 when R=1.4, B=1.3

  AWB vs. FWB R1.4B1.3
  slope R2
red 1.36 89%
green 0.99 91%
blue

1.13

 57%

 

DOY 159-307 (present) when R=1.8, B=1.8

  AWB vs. FWB R1.8B1.8
  slope R2
red 1.08 81%
green 1.13 79%
blue 0.63 8%

 

Timeseries plot

The timeseries plots show the result of rescaling the FWB photos to get the final GCC to approximately match the GCC from the AWB photos.

Scaling factors used FWB R1.4B1.3 FWB R1.8B1.8
  RB RGB RB RGB
Red 1.29 1.36 n/a 1.08
Green  n/a 0.99 n/a 1.13
Blue 1.38

1.13

 n/a 0.63
Method to calculating rescaling factors Joe taking photos outside his window and testing which scaling factors made the photo closest to what his eye could see

Linear regression

 Did not try since RGB scaling factors worked better for MB Linear regression
Conclusion The result from the linear regression "Rescaled RGB" seems closer than the results from Joe's comparison "Rescaled RB". Could rescale these photos by RGB scaling factors. Could rescale these photos by RGB scaling factors

 

After all of this rescaling analysis, Joe and I decided it wouldn't be worth the trouble of processing the FWB photos for our own database. FWB photos don't provide a significant improvement over AWB photos, especially if we use moving averages and take the 90th percentile of our GCC data anyways.

 

Mayberry data in database:

2015-08-19 through 2021-10-23 (DOY 296): Canon PowerShot A480

2021-12-01 (DOY 335) to present: Picam AWB photos

Mayberry photos available:

2021-12-01 (DOY 335) to 2022-06-09 10:15 (DOY 159): FWB photos rescaled using R(1.36), G(0.99), B(1.13)

2022-06-09 10:45 (DOY 160) to present: FWB photos rescaled using R(1.08), G(1.13), B(0.63)