beyond-cinema-stories
USC Institute of Creative Technologies KOMODO Testing Results
November 2nd, 2023

USC 3D Multiview

Multiview Stereo cameras are used in computer vision to generate 3D geometry. In order to get the best output of the 3D mesh/ maps we require cameras which have big pixel pitch, high resolution, high Dynamic range, great Signal-to-Noise Ratio (SNR), Linear output and fast recording framerate. In this test we used the RED Komodo cameras synchronized to the Light Stage by using Genlock and Timecode to capture the color chart with different exposures as well as different settings to test the camera parameters. We push the camera sensor to Overexpose and Underexpose to see the Linearity range of the camera. We test the SNR differences between the Recording qualities of the camera. We show that the RED Komodo has an excellent Dynamic range, great SNR and fast framerate with high resolution which outperform the previous still DSLR cameras we were using.

The Experiment setup:

A Black KOMODO with the custom ICT firmware where the Timecode was fixed and attached to the Canon 85mm f2.8 RF lens was used for this experiment. The expansion adapter and the KOMODO Link adapter were attached for control and data accessibility. The project settings were 6K 17:9 @ 30FPS ISO was set @ 500 in HQ setting since that’s the maximum resolution and framerate we can synchronize the cameras at, and the F-stop was f11. We calibrated the camera with the above settings before shooting. The exposure values were used in Time rather than angles and the values used were multiples of 2:

1/30, 1/60, 1/125, 1/500, 1/1000, 1/2000, 1/4000, 1/8000

The camera is placed just outside the Light Stage on an Aluminum gantry at a distance of 1.35-1.4m away from Color chart at the middle of the stage which is perpendicular to it. The stage is then illuminated with a flat lighting condition, the Pulse-Width-Modulation (PWM) of the LEDs are set at Max so that they don’t flicker at all during this experiment.

The List of Experiments:

  1. Signal-to-Noise ratio under flat illumination for Linearity
  2. Signal-to-Noise ratio for underexposed
  3. Signal-to-Noise ratio for overexposed
  4. Signal-to-Noise ratio for MQ and LQ setting.

EXP1 : Signal-to-Noise ratio under flat illumination

The Stage is set at Analog Intensity of 100 (controls current intake for our application specific light boards) and the PWM is set to max so that there is no flickering of LEDs during the shoot. The camera was focused on the color chart after calibration. (It’s in manual focus and not in Auto-focus mode since we wouldn’t want to have focus shifting during a test). Then Light Stage Master was used to trigger the cameras using GPI and then sample their framerate using GPO and then trigger the Light stage in accordance with the GPO of the camera and captured 4 frames for each exposure setting (we prefer to take at least 2-3 frames since we account for camera inconsistency if any).

The 9 exposure settings are captured in .R3D format then they are converted to EXR with no compression using REDline.exe and then the pixel values are sampled from the white square of the color chart (90% reflective).

The ploted results are as shown in the figures below:

Figure 1: Signal-to-Noise ratio under flat illumination

The curves are pretty good, they are not the smoothest we have seen but since RED KOMODO is a video camera focusing on the speed of capture its understandable. The Green spectrum seems to be more linear than the RED and the BLUE spectrum of the capture. The SNR values are much higher than the cameras we have tested before like the 1DX and the Ximea cameras which is what we were hoping for.

We also tested the pixel values in the Color Chart of the Greyscale line to see how linear the camera response is when compared to the reflectivity’s of the greyscale values on the chart.

Here are the results:

Figure 2: Signal-to-Noise ratio greyscale

As you can see from the charts the output values are very Linear.

EXP2 : Signal-to-Noise ratio for Underexposed

The Stage is set at Analog Intensity of 25 (controls current intake for our application specific light boards) and the PWM is set to max so that there is no flickering of LEDs during the shoot. The camera was focused on the color chart after calibration. (It’s in manual focus and not in Auto-focus mode since we wouldn’t want to have focus shifting during a test). Then Light Stage Master was used to trigger the cameras using GPI and then sample their framerate using GPO and then trigger the Light stage in accordance with the GPO of the camera and captured 4 frames for each exposure setting (we prefer to take at least 2-3 frames since we account for camera inconsistency if any).

The 9 exposure settings are captured in .R3D then they are converted to EXR with no compression using REDline.exe and then the pixel values are sampled from the white square of the color chart (90% reflective).

The ploted results are as shown in the figures below:

Figure 3: Signal-to-Noise ratio underexposed

In these graphs if you zoom I into the origin of the graphs there are points which are not exactly sitting on the curves, we just wanted to verify if these points, for example the Blue channel zoomed in looks like this:

You can see that the 2nd and 3rd points are almost the same value, from this we have less confidence in the linearity of the pixel values we get the from bottom 10% of the image’s underexposed range. Since it sometimes can’t differentiate properly between those 2 exposures.

EXP3 : Signal-to-Noise ratio for Overexposed

The Stage is set at Analog Intensity of 200 (controls current intake for our application specific light boards) and the PWM is set to max so that there is no flickering of LEDs during the shoot. The camera was focused on the color chart after calibration. (It’s in manual focus and not in Auto-focus mode since we wouldn’t want to have focus shifting during a test). Then Light Stage Master was used to trigger the cameras using GPI and then sample their framerate using GPO and then trigger the Light stage in accordance with the GPO of the camera and captured 4 frames for each exposure setting (we prefer to take at least 2-3 frames since we account for camera inconsistency if any).

The 9 exposure settings are captured in .R3D then they are converted to EXR with no compression using REDline.exe and then the pixel values are sampled from the white square of the color chart (90% reflective).

The ploted results are as shown in the figures below:

Figure 4: Signal-to-Noise ratio overexposed

As we can see from the graphs when we go to the overexposure region of the data the curve structure falls and the SNR values increase dramatically, we can’t use this inconsistency while we run our software pipelines hence we don’t use the top 10-15% of the image exposure range since it’s not reliably linear.

EXP4 : Signal-to-Noise ratio for MQ and LQ setting.

The Stage is set at Analog Intensity of 100 (controls current intake for our application specific light boards) and the PWM is set to max so that there is no flickering of LEDs during the shoot. The camera was focused on the color chart after calibration. (It’s in manual focus and not in Auto-focus mode since we wouldn’t want to have focus shifting during a test). Then Light Stage Master was used to trigger the cameras using GPI and then sample their framerate using GPO and then trigger the Light stage in accordance with the GPO of the camera and captured 4 frames for each exposure setting (we prefer to take at least 2-3 frames since we account for camera inconsistency if any).

The 9 exposure settings are captured in .R3D then they are converted to EXR with no compression using REDline.exe and then the pixel values are sampled from the white square of the color chart (90% reflective).

The ploted results are as shown in the figures below:

Figure 5: Signal-to-Noise ratio for MQ and LQ settings

This last experiment was just to see how the MQ and LQ settings of the R3D recording holds up to the SNR of the HQ setting as we can see the MQ comes close to the SNR of the HQ especially in the lower end of the image range and higher range of image range, but the LQ falls short right after the initial low image range. So, for our high-quality scans we only recommend HQ setting but if we want to capture long performances then we can even try using MQ is the memory is not enough for HQ else we recommend shooting in HQ to always get the best results.

The Ideal Camera we need for the Light Stage to get the best data with least time constraint/ effort constraint:

  1. A sensor with good pixel pitch (so that we can get good SNR values/ low noise)
  2. High megapixel count (so that we have a lot of resolution for the parts of the face)
  3. Fast frame rate (less motion blur by the subjects’ involuntary movements)
  4. Good Dynamic range (so that we can capture the One Light At a Time (OLAT) with great efficiency)
  5. Easy download of data (directly to a connected PC or server)
  6. Global shutter (to prevent Rolling shutter artifacts since we are also using PWM for the LEDs)
  7. Easy interface of the camera for control and focus
  8. Easy multi-camera synchronization (for the Multi View Stereo pipeline and for view variance)
  9. Small size for easy mounting around the Light Stage

The RED KOMODO satisfies almost all the points which are noted above maybe bordering on point 1 & 2. Its sensor response curves are great, and its Dynamic range is much better than the cameras we have tested. We are also getting great results from it in the preliminary scans which we have done. In the market it’s the best camera for this purpose.