change ISO vs increase RAW exposure

in terms of digital noise:
is it the same to increase the iso in the camera before taking the
picture that take the photo in RAW with the lowest iso and then
increase the exposure with the RAW-reader program?

 

In message <>,

Unless the camera's analog-to-digital converter is really bad, it is
always better to saturate the range of RAW values at a higher ISO than
to under-expose at a lower ISO. Under-exposure should be avoided
whenever possible, as it causes posterization of the signal and the
noise. I only under-expose when there just isn't enough light, such as
shooting at the highest ISO at night.

A well-exposed image at ISO 1600 looks orders of magnitude better than
the same shot at ISO 100, relatively under-exposed by four stops (same
aperture and shutter speed).

This is very important to understand if you want maximum quality
captures. When the contrast of the scene is low, crank that ISO, and
boost the exposure compensation, instead of exposing in the middle zone
at ISO 100. Just watch out that you don't clip important detail in the
RAW capture.
--

 

Normally no but:

Depends on the camera and the max designed ISO of the sensor. If the camera
is "boosting" the sensor to get higher ISO setting, then a lower ISO
underexposed a stop then increasing the exposure in RAW can get a cleaner
image. An example of what I mean:

On my olympus E300 (shooting RAW): a native ISO 800 shot is cleaner than a
ISO 400 one underexposed 1 stop and then increasing the exposure a stop in
developing to end up at the same place. On the other hand it does produce
cleaner results shooting at ISO 800 1 under and pushing the exposure a stop
in development than shooting at native ISO 1600.

Only testing for yourself will show you if this works for your camera/sensor
and is probably only worth doing at the highest ISO's.

 

Interesting! Are you suggesting that for a normally lit scene (ie. no
difficult exposure issues) and all else being equal, that it's best to
shoot at 1600 vs. 100; unlike the film world?

Can you elaborate or site a source I could read?

 

You could start by reading the message you quoted more carefully...

David J. Littleboy
Tokyo, Japan

 

In message <>,

What I am suggesting is that for any given combination of aperture and
shutter speed, you will get a cleaner image with the highest ISO that
doesn't blow out the RAW data. For example, if you find yourself
shooting in a dense fog, and the aperture and shutter speed are at
usable limits at ISO 400, you will get a better capture at ISO 1600 and
+2 EC, than you will at ISO 400 with 0 EC, using the exact same shutter
speed and aperture. The exposure on the sensor is *exactly* the same
for each; except for determining automatic exposure and manual metering,
the ISO has no affect on exposure of the sensor. Its only affect is the
scaling of the analog-to-digital conversion.

I've tested this in many scenarios on both the 20D and 10D, and with the
same absolute exposure (aperture and shutter speed) the higher ISO image
is always as good or better than the low-ISO shot; never worse.

Some people understand intuitively that grossly under-exposing at ISO
100 is to be avoided, but few people realize that a low-contrast scene
with moderate light can also benefit from going to a higher ISO, with +
EC. Film thinking still rules the day. Both film and digital get
grainy/noisy from under-exposure. Only film gets low-contrast (and
therefore, grainy upon restoration) from over-exposure. Digital records
better right up to the point before it clips.

Of course, if you allow the exposure to vary automatically in the
different ISOs, then the lowest ISOs will be cleanest (but may sacrifice
DOF or action-stopping to shoot at that ISO).
--

 

Short answer: you would never want to shift from ISO100 to ISO1600
on a shot that you can expose correctly at a decent shutter speed at
your desired aperture, but you might often want to shift from 100 to
200, or from 200 to 400 to deliberately overexpose slightly (or avoid
mild underexposure), and then correct that in post-production.

For the longer answer, you have to understand the two main issues
involved.

First, the amount of bit depth allocated to storing the values in
an image is not linearly spread across the entire brightness range.
There is three times as much information available at the upper two
stops than there is in the lower three, given a camera with a 5-stop
dynamic range. You might want to read "Expose to the Right" and
"Understanding Histograms" by Michael Reichmann
(http://www.luminous-landscape.com/tutorials/expose-right.shtml and
http://www.luminous-landscape.com/tutorials/understanding-series/understanding-histograms.shtml).

Second, there is the issue of how much noise is introduced at each
ISO level, and how much blur is introduced by a slower shutter speed
in a handheld shot or with a less than perfectly stable tripod (for
very long exposures).

If you know in advance that you are going to post-process from the
raw image, you want to avoid underexposure when possible, and even
slightly overexpose in some cases, as long as you don't blow your
highlights. When possible, get extra light by going to a slower
shutter speed. If you are shooting by hand, and have run up against
your minimum speed for a sharp image, then you get to consider the
quality of your sensor.

A Canon EOS 20D has excellent progression up to ISO1600, but then
a sharp degradation at "H" or ISO3200. When I shoot, I'll bump the
ISO up as far as 1600 to avoid even slight underexposure or need to
take a faster shot (for instance because I need extra depth of field
by decreasing aperture or because the subject is moving), and will go
from 100 to about 400 if I want to put more data in the upper light
zones even if the shot is exposed decently. I will go to 3200 if I'm
losing the shadows (I take a shot, and the histogram shows that I'm
pressed up against the left hand side), but I try to avoid it except
in desperate cases. Adding noise is almost always superior on a
decent camera to completely losing information.

There's something of an extreme example of this for when I did a
review of night shots with the 20D a little while ago at
http://www.resonant.org/node/471 (check for the image that is dead
black at ISO400, and compare how much I was able to retrieve with
level adjustment, then compare again to the ISO3200 version, and the
level adjusted version of that).

Whether or not going to a higher ISO to get the extra information
when a shot is only slightly underexposed or exposed correctly but
with a lot of dynamic range to spare on the bright side requires a bit
of personal judgement. My recommendation, given that storage is
becoming cheap, is to do both whenever you're in doubt. Take a shot
at the lowest ISO where you are not significantly underexposed for the
shutter speed and aperture that you want. Put both shots through
Noise Ninja, adjust levels, curves, and/or shadows/highlights, and
then compare the results. After a while of doing this, you should get
a feel for when it's a good idea.

That's a decent general rule of thumb, actually: when in doubt,
take lots of pictures.

 

In message <>,

There is no ISO 3200 mode, per se, on the 20D. It is using ISO 1600,
but exposing for 3200, and doubling the digitized RAW numbers ... they
are all even except for half of the bad pixels, which are interpolated
to odd numbers. They are aboout 99.9% even.

ISO 3200 on the 20D is really a JPEG mode, as shooting RAW at 1600 and
-1 EC captures everything RAW 3200 does, plus it gives you one more stop
of highlights. For JPEGs, the story is different, as the extra
highlights are clipped away, anyway, and it is better to have the image
normally-levelled before conversion to 8-bit JPEG.

This is probably true of most if not all current digitals that start at
ISO 100 and go to 3200. The 10D's "1600" is 800, and its "3200" is
1600, done the same way. The difference is that on the 10D Canon
appears to be trying to fool histograms and RAW peepers by striping the
RAW data with solid and dotted horizontal lines that are all offset to
odd numbers, but the bottom line is that any position in the RAW bitmap
can only be one of 2048 odd or even numbers.
--

 

G Duran writes ...

I've tested this on a couple of cameras and I have to disagree with
those who are seeing "orders of magnitude" improvement. I'm seeing
roughly equivalent noise when I do this. Keep in mind that if you
increase ISO and then add exposure compensation you'll have a lot more
editing in the RAW converter than if you simply shoot at the correct
exposure at lower ISO. The other thing to worry about is clipping
highlights, which is much easier to do when you're working close to the
right edge of the histogram.

This is based on what I see with my cameras, you might get different
results with a different camera or a different RAW converter. So I
suggest you run some tests with your gear and decide for yourself if
it's worth the hassle. Here are results from such a test from one of
my cameras ...

First, I shoot a neutral subject (narrow tonal range) at each ISO at
the metered reading, then with -2, -3, + 2, and +3 stops compensation
(the cameras I use only go +/- 3 stops unless you expose manually).
Here's a screen grab of what these RAW thumbs look like in the RAW
converter before any RAW processing is done ...
http://members.aol.com/bhilton665/thumbs.jpg

This next link shows noise samples from four files, ISO 800 metered and
unchanged during RAW conversion, 800 underexposed by two stops and then
adjusted during RAW conversion for an equivalent exposure, 800
overexposed by 2 and adjusted down by 2 during RAW conversion for an
equivalent exposure, and ISO 200 as metered. The three ISO 800 shots
show the benefits of "exposing right" since the +2 exposure with -2 RAW
conversion has much less noise. But if you compare this +2/-2 @ 800
crop to the iso 200 crop you won't see much difference in noise. These
are all processed with the default setting in my RAW converter, 100
pixel crops from the center of the file and then blown up to 400% to
show any noise. At 100% (web view) any difference between 200 and 800
@ +2/-2 is neglible. I see basically the same trend comparing iso 100
metered to 400 +2/-2 and 400 metered to 1600 +2/-2.
http://members.aol.com/bhilton665/iso800_200.jpg

One more link ... this one shows iso 800 +3 exposure with -3 RAW
(actually -2.2 ec to get there). Again, there is little discernible
difference in the noise. I also included the histograms and as you can
see there is little room for error with the +3 shot since the histogram
is heavily biased to the right. This is shooting a gray card, most
"real life" scenes would have more contrast and would likely clip the
highs with +3 compensation.
http://members.aol.com/bhilton665/iso800_100.jpg

These tests used a professional camera with pretty low noise levels
(Canon 1D Mark II) and a very good RAW converter (Capture One Pro).
With smaller, more densely packed sensors and different converters you
might get different results.

Bill

 

John Sheehy writes ...

I don't believe this, based on what I've seen with Canon dSLRs. Can
you post an example showing "orders of magnitude" improvement (what
kind of camera are you using that's noisy at ISO 100?). Please name
the camera and the RAW converter you're using and explain how you are
doing this test too.

Thanks.

Bill

 

In message <>,

I don't know about anyone claiming "OoM" differences, except in the
realm of extreme under-exposure (like the 1600 vs 100 @ -4EC sample I
showed). Comparing lower ISOs, especially normal exposure vs high, is
less drastic.

Your experiments are only measuring noise. They don't have any
opportunity to be affected by signal quantization, which is finer when
you shoot at higher ISOs and the absolute sensor exposure remains the
same.

This is true, and this is one reason why experiments like yours are
difficult; you are using tools that assume that the brightest highlights
are specular highlights, and that no-one would "overexpose" on purpose,
which is totally wrong. Unfortunately, current RAW converters are
insufficient; they should all have tonal transfer curves, etc, but they
should also have pure linear pre-gain, and most don't. ACR, for
example, will take RAW values above a certain point and nail them to 255
in the output, no matter how much -EC you use.

When I shoot RAW, I often envision better software down the road. Much
of the noise in some cameras is patterned, like the bands in some dark
20D pictures. Most if not all of the current RAW converters cause
banding, because the readout offsets are different for each line, and
they use a global offset. I've already seen the patterns disappear
completely in DCRAW doctored to offset each line individually; the
problem is most people don't want to use DCRAW; they want the GUI
features of ACR and Capture One, which still use global blackpoint
offsets.

Well, part of the skill in "exposing to the right" is knowing how much
you can. You have to assess the scene, and decide how much.
--

 

In message <>,

You misunderstand ... Sensor and readout noises are not the issue with
very low ISO 100 exposures. The issue is *Posterization*, which can
distort the image much more than sensor/readout noise.

I don't trust the converters; I look directly at the RAW data. Canon
10D and 20D are the cameras.

--

 

Perhaps you could can provide us with a little background
here. Where did you come up with all of this information?
Did you get it from Canon?

-- Yip

 

In message <zto%d.26288$>,

This info is available by looking directly at the RAW data. A 10D ISO
"1600" or "3200" shot has RAW data that looks like this:

EEEEEEEEEEOEEEEEEEEEEEEEEEEEEEEEEEEOEEEEEEEEEE
EEEEEEEEEEEEEEEEEEEEEEEOEEEEEEEEEEEEEEEEEEEEEE
EOEOEOEOEOEOEOEOOOEOEOEOEOEOEOEOEOEOEOEOEOEOEO
EOEOEOEOEOEOEOEOEOEOEOEOEOEOEOEOEOEEEOEOEOEOEO
EOEOEOEOEOEOEOEOEOEOEOEOEOEOEOOOEOEOEOEOEOEOEO
OOOOOOOOOOOOOOOOOOOEOOOOOOOOOOOOOOOOOOOOOOOOOO
OOOOOOOEOOOOOOOOOOOOOOOOEOOOOEOOOOOOOOOOOOOOOO
EOEOEOEOEOEOOOEOEOEOEOEOEOEOEOEEEOEOEOEOEOEOEO

Here's an example from a real image, where the bitmap only shows the
least significant bit (the one that determines odd or even):

http://www.pbase.com/jps_photo/image/38841732/original

The 20D is like this at ISO "3200":

EEEEEEEEEEEEEOEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEE
EEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEOEEEE
EEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEE
EEEEEEEEOEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEE
EEEEEEEEEEEEEEEEEEEEEEOEEEEEEEEEEOEEEEEEEEEEEEEEEEE
EEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEE
EEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEE
EEEEEEEEEEEEEEEEOEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEE
EEEEEEEEEEEEEEEEEEEEEEEEOEEEEEEEEEEEEEEEEEEEEEEEEEE
EEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEE

What possible logical conclusion can one come too, except that these
numbers were originally one bit less?
--

 

The 20D programmers were chanting at the time?

[ducks back under table]

Jeff

 

This is really, really interesting. I'll have to play with this a
little. ISO3200 is visibly noisier than ISO3200 level-adjusted from
RAW, though, and I don't know why that would be if it's basically the
same mode. If anything, the interpolation should smooth things, no?

You mean a +1 EC? Or you mean playing with the EC in
post-production by one stop? If you mean going for +1 EC on the
camera, the extra light has to come from somewhere, either from
aperture or shutter speed, and you wouldn't be playing with even
ISO1600 unless those options were already blocked off. In that case,
your option is between shooting ISO1600 one or more shots underexposed
(thus wasting the extra RAW data space in the highest stop), or
forcing interpolation in the camera to move data out of the lower
stops.

This is true, but when half of those numbers are in the top stop
alone, you want to force your data over there if possible.

 

In message <>,

I'm not sure what you're talking about here, or if you understood what I
wrote. Let me restate it ...

When the 20D is set to "ISO 3200" the metering system works as expected;
it meters for ISO 3200. When the data is read off of the sensor,
however, the same gain is applied as is applied for ISO 1600, and then
the numbers are doubled after digitization. 0 stays 0, 1 becomes 2, 2
becomes 4, 3 becomes 6, ... 2047 becomes 4094, and anything higher than
2048 becomes 4095. Also, any mapped out pixels, interpolated from
nearby pixels, have a 50% chance of being odd, as they are interpolated
after the doubling.

I mean -1 EC, and ISO 1600, in the camera. The midtones and shadows
need +1 EC in the RAW converter, or post-processing The highlights can
be clipped or squished at will.

No, that wasn't what I was talking about. Setting the 20D to ISO 1600
and +1 EC is basically "shooting at" ISO 800, with the top of the
highlights clipped one stop lower than if the camera were set to ISO 800
and ) EC, but with one bit more depth in the range captured; IOW, it's
like having ISO 800, but instead of 0 to 4095, having 0, 0.5, 1, 1.5,
.... 2047, 2047.5.

That's called shooting at ISO 1600 and 0 EC. You shoot at ISO 3200, or
1600 with -1 EC, because of lower light (or the need to stop action, or
increase DOF).
--

 

In message <>,

Unfortunately, most camera manufacturers don't seem to care if you know
how the RAW data is doing, even though that's really the most important
aspect of exposure (quality-wise) at any given ISO. The only way I know
is to get a program like IRIS that lets you look directly at the RAW
data, and see what RAW values are present and compare them to the
histogram and/or the JPEG, so you can get an idea of how much headroom
there is. The only cameras I know well in this regard are the Canon 10D
and 20D; the 10D starts to blink or clip the histogram when the green
channel is about 1 stop from the green RAW highlight limit. The 20D is
more like a 1/2 stop, as it stuffs an extra 1/2 stop into the top of the
8-bit output (about 245 to 255).
--

 

Well, partly, I meant to write "visibly noisier than ISO1600", and
partly, I may just be very confused.

I'm with you so far. What I'm trying to work out in my head is
where the extra noise is coming from at ISO3200 (I thought it was
related to gain), and whether or not there's any advantage in terms of
dynamic range to shoot at ISO3200 than at ISO1600 (if I'm
understanding you correctly, it seems that the answer is no for RAW,
and the only time you'd ever want to use ISO3200 is when you're
shooting straight to JPG).

Okay, either I have a sign backwards, or I'm not at all
understanding what you're doing here. Just to be absolutely clear,
when you say "-1 EC", you mean "underexpose by one stop" (relative to
what the camera light meter thinks is appropriate exposure), correct?
This combined with ISO1600 is two stops darker than you would get by
default with ISO3200, squishing everything far to the left in a scene
that is dark to begin with. I'm not understanding why you would want
to do this.

If you mean "overexpose by one stop", then this makes more sense,
but the sign seems to be contrary to what I see in the camera and in
the EXIF data in the saved files.

 

In message <>,

We're on different trains of thought.

Noise comes mainly from the signal-to-noise ratio. The noise in the
sensor is the same at all ISOs. The main reason that the same scene
auto-exposed at ISO 1600 is noisier than the same scene at ISO 100 is
that the signal is a lot weaker at ISO 1600, and the noise is
consequently higher in the ratio. There is also some extra noise and
distortion from the amplifier, that becomes a bit of a factor at the
highest ISOs.

Well, another reason is if you want to shoot 1600 with auto-exposure,
but three stops under, the only way to emulate it in auto mode is with
3200 and -2 EV, since the 20D only has EC down to -2.

It's one stop darker, in the JPEG, or default RAW conversion. I don't
know where you're getting the "2" from. In the actual sensor exposure,
they are exactly the same (but the digitization will only digitize half
the dynamic range with "ISO 3200" that it does with ISO 1600 and -1 EC).

Because you don't have a choice, a lot of times. Even if you set the
camera to ISO 1600 and PLUS 2 EC in shutter priority mode, if there
isn't enough light the exposure is going to be negative anyway, even
thought the camera wasn't set that way, Many fast lenses also have
horrible optics wide open, such as the Canon 50mm f/1.4 and 24mm f/1.4L.
Just stopping down to f/2 can cause a drastic improvement in corner
optics, so setting the camera to shoot a little on the dark side to
avoid 1.4 as much as possible is an option.

--