Sunday, July 6, 2014

Consolidation Series: Part 2



Following on from the first part of this ‘Consolidation Series’ of posts on ML-based workflows, here is another case study that hopefully will help convince the ML sceptics that they should take the ML plunge.
 
Case Study #2: Extending the dynamic range (DR) of a single image: 

  • We all know our (expensive) DSLR EOS cameras are limited by the ‘laws of physics’ and some settings that the manufacturers choose to make in the firmware. At ISO 100, my 5DIII is said to have a DR of just under 11 stops; at ISO 1600 this drops to just under 10 stops.
  • Without ML and being confronted by a high DR scene, one ‘recovery’ strategy is to take exposure brackets, ie varying the shutter speed, and then, in post, blend them or tone-map them together. Done right, the resultant images can be made to look natural; done ‘wrong’ they can be made to look ‘radioactive’, ie have a so-called HDR-look.
  • There are two problems with ‘normal’ bracketing, both related to speed. First the shutter speed of the resultant brackets may prohibit handholding. Second the subject’s speed (motion) may create (ghost) artifacts when the resultant brackets are merged. 
  • With ML we have another tool: it’s called Dual-ISO. As most are aware our sensors are composed of thousands of ‘sensor sites’ and these are clustered into groups of four, to create a so-called Bayer pattern. Thus the colour at a given pixel is estimated by sampling the RGB values in the near neighbourhood of the pixel. The Bayer pattern is based on human visual response and laid out in an RGGB manner, like this:

RG RG RG RG RG RG ...                  All set to ISO X ‘gain’
GB GB GB GB GB GB ...
RG RG RG RG RG RG ...
GB GB GB GB GB GB ...

  • Those gurus at ML (really Alex) worked out that certain Canon cameras, eg the 5DIII, can be made to alternate the ISO every third line. In other words the sensor will capture the data on the first two (Bayer) lines at, say, ISO 100 and the next two (Bayer) lines at, say, ISO 1600, ie like this:

RG RG RG RG RG RG ...                  ISO 100
GB GB GB GB GB GB ...
RG RG RG RG RG RG ...                  ISO 1600
GB GB GB GB GB GB ...
RG RG RG RG RG RG ...                  ISO 100
GB GB GB GB GB GB ...
...           

  • Remember that the ISO setting we use simply configures the analogue gain that is applied to the sensor read voltage (measuring the captured photoelectrons), before analogue-to-digital conversion is carried out. Thus a half well of photons/electrons at ISO 200 ‘looks’ like a full well’s worth to the AD converter, ie the voltage at that senor (R or G or B) site has been boosted by a factor of 2. Similarly, at ISO 1600, we only need 1/16 (1/(2x2x2x2))of the photons in the well to ‘look’ like it is full up. Which gives us the ability to capture low light scenes which look as if they are metered ‘correctly’, albeit with a resultant ‘boost’ in noise as well and a reduction in DR. 
  • To make use of the Dual-ISO we simply select the ‘boost or gain’ we wish to give the image, say, selecting a boost ISO of 800, or 3Ev. Unprocessed the captured image will look strange, ie the histogram will be strongly biased to the right; and if you zoom in you will clearly be able to see the alternating (ISO) scan lines. For example:


Dual-ISO w/o post processing
Details of alternating 100/800 ISO capture lines
  • However, if you throw your Dual-ISO image at the ML supplied ‘correction tool’, called CR2HDR.exe, either in command line mode or using the Lightroom CR2HDR plugin, magically you will have a 16-bit TIF that has the benefits of both the ISO 100 and ISO 800 image, all compressed into a single image. The downside being some loss in vertical resolution, but not 50% as you might expect. For example:


Post processed with CR2HDR.exe
  • Thus, one way of looking at Dual-ISO is to consider it a two-bracket capture strategy, but accomplished by taking a single in-camera image, ie no motion-induced artifacts between the two brackets. Both brackets are taken at the same shutter speed and at the same time.
  • In the case of the 5DIII, the Dual-ISO approach provides a boost to the base ISO 100 DR of about 3Ev. Pushing the boost ISO to 1600 and beyond can be done, but you will begin to see image ‘degradation’. I personally tend to use a boost ISO of 800, or, at a push, 1600. 
  • In addition to simply increasing the DR capture ability, Dual-ISO may be seen as a form of post-process recoverable fill-flash. In other words, in post shadows, say on a face, may be addressed as we have boosted the base image by 3Ev, ie the data is there to be used. 
  • It is worth noting that Dual-ISO can also be applied to RAW video; more on this in a subsequent post. 
  • Finally, with reference to the first post in this consolidation series, ETTR and Dual-ISO can be statically and dynamically combined, such that the auto-ETTR can decide to call the Dual-ISO if the scene needs it or be on all the time. In addition, ML provides some useful workflow choices. For example, you can elect to tell the camera to alternate its Dual-ISO usage, thus ending up with a non-Dual-ISO and a Dual-ISO image, at the same shutter speed but at (fractionally) different times. Thus giving you a non-Dual-ISO ‘fallback’.

For me, ETTR and Dual-ISO represent two of the best features that ML provides. In subsequent posts I will continue to look at other case studies where an ML-based workflow can help with your photography.

No comments:

Post a Comment