Tuesday, February 6, 2018

Optimised Strategy for hand held Exposure Bracketing

As photographers we are very aware that, although our digital camera-systems are impressive, they still need to be managed intelligently: technology has yet to reach the state where we can ignore, say, dynamic range (DR) or (shadow) noise.

Although we can recover many of the inherent weaknesses in our camera-systems, eg through exposure bracketing, these techniques are not always usable; for example: shooting in an area that prohibits tripods; trying to shoot on a pitching ship; or trying to bracket on a very windy day. Thus, there is still a need try and get the best out of a single image and/or use a more optimised approach to handheld bracketing.

As Canon Magic Lantern users know, the Dual-ISO feature is a handy way of ‘boosting’ a single image’s DR by up to about 3Ev. But this DR-boost only brings the Canon DR more in line with other camera manufacturers, who offer higher DR cameras than Canon at low ISO. The chart below illustrates where Canon’s DR lies relative to Nikon’s and Sony’s cameras.

Also, as we know, Dual-ISO comes at a cost, ie resolution.

Thus, for some time I have been looking for a ‘better’ High DR bracketing approach when forced to handhold, especially in very high DR environments, ie beyond the ML Dual-ISO capability.

Although I have used information from many sources, eg clarkvision.com and photonstophotos.net, the defining work that I have used to inform this post comes from MIT and the work presented by Hasinoff et al: https://people.csail.mit.edu/hasinoff/hdrnoise/

Also, usually hidden from simple users like me, I have also looked at the International Standards that camera manufacturers use, eg ISO 12232:2006 – Photography – digital still cameras – Determination of exposure index, ISO speed ratings, standard output sensitivity and recommended exposure index. I have found reading these to be educational.

For example, clarkvision.com summarises the ISO characteristics of my 5D3 as follows:

I have estimated the approximate (max) total noise as SQRT[Max Signal + RN*RN]. Note that the DR vs ISO curve shows the 5D3, like all other Canon EOS cameras, is not ISO invariant over the full ISO range, but, in the case of my 5D3, is above about ISO 3200. Thus, shooting above this is not likely to be of any real advantage (unless I needed to use such a high ISO to control subject motion).

The hand holding bracketing ‘use case’, I’m talking about in this post, is a form of ‘time limited photography’, ie where the total exposure is limited in some way: in this case because we are hand holding.

For example, let’s assume we are shooting wide, ie focal length less than, say, 50mm on a full frame, and I haven’t had a drink [:-)], and thus I set a single image hand holding limit of, say, 1/30 sec: this could be lower or higher, but 1/30s is a reasonable number to use.

Normally, when exposure bracketing, ie we are ignoring Dual-ISO for now, we seek to find a bracketing scheme that covers the dynamic range of interest, ie the shadows that are of interest with sufficient detail, to the highlights of interest that are not clipped.

If we have the luxury of time, we would seek to intelligently survey the scene, eg with an exposure meter, and work out an optimum bracketing strategy. Or, if we were on a tripod, we could make use of the Magic Lantern Auto Bracketing.

But we are hand holding, thus our slowest shutter speed needs to meet our hand holding limit, ie 1/30s in this example.

If we follow a typical bracketing model, knowing our slowest shutter speed is 1/30s, we would use our lowest ISO, ie ISO 100, and take three brackets at, say, 1/30s, 1/120s, 1/480s, ie at differences of 2Ev and 4Ev from the base. But, of course, these shutter speeds ignore the needs of the scene DR, eg daylight vs dusk; external vs internal. They only illustrate the hand holding problem.

In general a reasonable (scene-informed) bracketing strategy is to meter for the highlights and bracket for the shadows. In this post we won’t worry about how the highlight exposure is obtained, eg through Canon or ML (ETTR or Raw Spot) metering. We’ll assume a good base (highlight) exposure is set.

As we are trying to derive a robust strategy, let’s assume that our aperture is also fixed, eg for artistic reasons or depth of field control; thus our base (slowest) shutter is not going to be slower than 1/30s.

So where can we go to get our shadow-biased brackets?

The only place is, of course, to use ISO, ie higher ISOs than our base (highlight) exposure.

But I hear people saying that will not work: as high ISOs mean more noise.

But noise is complex and made up of many components, eg photon or shot noise, thermal or dark current noise and electronic or read noise etc. Plus noise is not the only metric we need to worry about in photography, there is also signal and signal to noise ratio (SNR).

For example, at low ISOs, (electron) read noise is higher (sic) than at high ISOs. Also, (electron) shot noise is a function of the SQRT(number of captured photons). Noting, for the same apparent exposure, eg the histogram reading, the number of photons at ISO 1600 is lower, by a factor 16. For example the 'full well' value for my 5D3 at ISO 100 is some 68900 electrons, and at ISO 1600 it is 4050.

Also, at least for my 5D3, and it will vary from camera to camera, the full well SNR at ISO 1600, compared to that at ISO 100, is less than a stop, eg 10.9 to 10.1 stops; as we are not in the camera’s ISO invariant zone.

To help ‘make the case’ for ISO bracketing here is a plot taken from the MIT work.

The key take away from the chart, for time critical bracketing, is that there is a change in bracketing strategy indicated when the total exposure time of the brackets, ie the sum of the shutter times in the bracket sequence, is, say, below a second or so: as it will be in our hand held use case.

For long total exposure times, say greater than a second or so, low ISOs are a better approach; but for short total exposure times then high ISOs are better.

For example, if we are hand holding, 3 brackets at 1/30s, the total exposure time will be about 0.1s; and we will typically not be seeking to bracket handheld much above, say, 6 brackets. That is we can consider a sensible (maximum) total exposure limit of, say, 0.2s for a hand held bracketing sequence.

If we look at the MIT results at, say, 0.2s and faster, we see a clear hint at an optimum bracketing strategy, ie use high ISO over low ISO when the total exposure time is low.

Thus ISO is not to be feared, especially if you are fighting other limits, as we are in the hand held bracketing use case. We need to embrace ISO and exploit it.

But how should we use it?

Based on some experimentation, as well as ‘intuition’, I suggest there are four possible schema:
  • we could see if we can find a fixed (higher) ISO, to ensure the slowest bracket remains above our hand holding limit, eg 1/30s here, and continue to vary shutter speed, but in dark scenes we will likely fail with this approach;
  • we could vary (ie bracket) ISO at a fixed shutter speed – but not all cameras allow ISO bracketing so you may need some addition technology, eg CAMRANGER or REMOTE PROMOTE;
  • we could optimally vary both shutter speed and ISO to maximise the SNR. This is the approach used in the MIT work, but this requires access to computer-based integer programming solutions;
  • Finally, if shooting Canon, we can exploit Magic Lantern’s auto bracketing with ISO shifting capability, which will ISO bracket to a maximum ISO, eg the start of the ISO invariant zone, then continue bracketing using shutter speed. But one ‘weakness’ of the current ML auto bracketing is that you can’t specify a minimum shutter speed, eg to cover the hand holding use case.
This conclusion, ie use high ISOs, may appear to some to be heresy. But let’s not fight the MIT data, let’s do a practical (Magic Lantern based) test: using the ML schema above as an illustration.

That is use ML ETTR to set the base exposure, ie to ensure we capture the highlights, using the slowest shutter speed you can get away with (1/30 in this example) at the lowest ISO you can. Note the faster your base shutter speed the more headroom you will have for (Tv) bracketing the shadows beyond the max ISO limit you set. Thus set the ML Auto Bracketing to Tv and switch ISO shifting on; having set the Canon AUTO ISO setting to, say, 3200 (in the case of my 5D3).

Here is the result of a test I just carried: Hand held indoors in the evening at 12mm and at F/5. Shot with a base exposure, for the highlights, of 1/80s at ISO 100. The other three images, calculated by the ML auto backeting, with a 3Ev delta and a four bracket request, were taken at: {ISO 800, 1/80}, {ISO 3200, 1/40s},{ISO 3200, 1/5s}. I recognise the 1/5s is marginal, but I have no way of limiting the minimum shutter speed.

Here are the four brackets as seen in Lightroom.

After some (non-optimised) LR processing, following HDR merge, this is what the final hand held capture looks like:

If you have got to this point, and are still reading: well done.

This hasn’t been an easy post to write, as I didn’t wish to regurgitate all the hard work that Hasinoff et al at MIT have done: 'just' use the insight it has created. I have thus only guessed at an optimum approach for hand held bracketing, ie without proof. However, I hope others explore the approach, ie use ISO bracketing, with or without combining it with shutter bracketing, eg using ML’s Auto Bracketing feature.

As usual I welcome feedback, including alternative views and corrections ;-)

No comments:

Post a Comment