The surprise comes from a little ISO test, which shows that ML can
improve dynamic range and highlight rolloff in movie mode (see
attached files). How much? Let's find the answer together by running a
little experiment.
Background
-----------------
There are two methods for obtaining in-between ISOs:
1) by asking the camera to set the ISO at 160 (for example);
2) by setting a full-stop ISO (let's say 200) and applying some
digital gain (let's say -0.3 EV) by changing DIGIC registers.
In these two cases, the resulting images will have:
- the same overall brightness in midtones and shadows;
- the same amount of noise;
- different highlight characteristics [!]
In the attached files, you can see that ML's ISO 640 (5D2, Neutral
0,-4,-4,0) retains more highlight detail and has a smoother transition
to white, compared to Canon's ISO 640.
So, this experiment will let us discover the best ISO settings for
difficult scenes (in particular, high-contrast ones, where you would
normally consider HDR video).
Test overview
--------------------
Under Debug menu you will find "ISO Tests", which can be divided in two:
* Real-world tests, where you record a scene with various settings, at
similar overall brightness; you will get a few test frames and tell me
which one(s) you like best and why.
* Synthetic tests, where ML computes the response curve at some fixed
ISOs; you will send me the results (log files) for analysis.
The result of these tests will depend on picture style and parameters
(especially contrast). Please test with any picture style you like.
I've ran a few tests with Neutral 0,-4,-4,0 and Flaat12p, but the data
is not sufficient to draw a conclusion yet.
How to run the real-world test
--------------------------------------------
This test will alter the settings while recording, to see what's best
for some particular scene. It takes roughly one minute.
- Find a contrasty test scene (one for which you would normally use
HDR video; that is, with bright highlights and dark shadows);
- Choose your favorite picture style;
- Enable movie logging;
- Expose the scene so you get both clipped highlights and clipped
shadows; use a full-stop ISO (100, 200..) and try not to choose
extreme shutter speeds;
- Reduce the bitrate (CBR 0.8 should be fine) - it's needed because ML
will save a few uncompressed still frames, and recording may stop;
- Choose "Debug -> ISO Tests -> Movie test" from ML menu.
After one minute, you will get two movies (one with HTP off and
another with HTP on); ML will alter ISO and shutter speed every two
seconds, and save a still frame (silent picture - YUV422) for each
setting combination. You will get 20 still frames (*.422 files).
Convert these pictures to JPG (see
http://magiclantern.wikia.com/wiki/Unified/UserGuide#silent-pictures
for a howto), post them here, together with the log file (or send
directly to me) and say which one(s) you like best, and why.
After hearing your thoughts, I'll tell you what settings I've used for
each frame (this should make the results as unbiased as possible).
How to run the synthetic test
--------------------------------------------
This test will compute the response curve at different ISOs. It takes
roughly one hour (maybe a bit more).
You will need:
- A Canon lens (kit lens and 50/1.8 are both fine);
- A controlled light setup with a white wall or screen. I've used a
second camera body to provide a white test screen, with the Flashlight
function (see attached picture). You can use a computer monitor too.
It's a good idea to avoid natural lighting (you can cover the setup
with a cardboard box, for example, or place it in a dark room).
- Strong batteries.
Before running the test, set the correct white balance for your test screen.
Under Debug -> ISO tests, you will run these 3 tests:
- For ISO 100 multiples (ful-stop ISOs), 160x (Canon method) and 80x
(ML method, with DIGIC ISO gain set to -0.3 EV);
- For ISO 70/65/50 multiples (ML method);
- For HTP, with ISO 200x (full-stops), 140x (ML method) and 100x (ML method);
To compute a response curve, ML will start from 1/30 seconds wide open
and will underexpose by changing shutter and aperture, in 1/8-stop
increments, over a total range of 16 stops. It takes roughly 1 minute
to compute one curve. If done correctly, you should get something like
in the attachment, starting with a flat section at the top and ending
with a flat section at the bottom.
Each test takes roughly 20 minutes. The result is a series of TXT
files which contain the data points. Zip all these files and send them
to me, mentioning what camera and picture style you have used.
Be careful if you use a second camera body to provide a white screen.
The camera will turn off after 30 minutes, even if you disable power
saving, so you need to re-activate the flashlight after each test.
----------------------------------------------
Back to so-called release notes. In this update you'll find these
important changes:
DIGIC-based refactoring:
- Silent pictures no longer use powersave tricks to pause LiveView;
- Display gain (night vision) and highlight recovery feature
(previously known as Highlight++) are now available under a single
control in the ISO submenu, called DIGIC ISO gain: positive values are
for night vision, negative values are for highlight recovery and low
ISOs.
- More ISOs with low noise: 60, 65, 70, 80, 85, 90 and their
multiples. Set a negative DIGIC ISO gain to enable them.
And a few handy shortcuts:
- Configurable arrow keys: besides ISO/Kelvin, you can also toggle
audio gain, shutter, aperture and others - check the Tweaks menu, the
shortcut key is camera-specific;
- Rack / follow focus: while recording, press MENU to save current
focus point and PLAY to return to it. No need to configure it from the
menu (just enable follow focus and start recording).
And also the hope that I didn't break anything :)
Happy Easter!
> --
> http://magiclantern.wikia.com/
>
> To post to this group, send email to ml-d...@googlegroups.com
> To unsubscribe from this group, send email to
> ml-devel+u...@googlegroups.com
> For more options, visit this group at
> http://groups.google.com/group/ml-devel?hl=en
advanced iso menu is very good anyway!
it's not a fair comparation, since you have no blacks (your levels starts at 18 out 255).
it's overexposed for this test cause it's all about highlights vs blacks.
you see perservation of highlight but you can't see loss in black because you don't have it at all.
1: the test is about improving the highlights, not shadows
2: its a real world test in that its a picturestyle i would use in my work. I have found that picturestyles that allow blacks to dip below a certain value contain lots of compression artefacts. Not everyone wants crushed blacks.
3: in the proper test, i could see a big difference in the shadow noise whilst using this pocture style, so even with the bottom end lifted, it still makes a difference.
4: As alex said, you are free to test with any picture style
5: the two images were shot with the same picture style so the comparison is completely fair.
Please post your own test results to show how me using this setting has in some way made the test unfair.
Perhaps your monitor is automatically adjusting or something, because I cannot see any difference in the shadows on the manga image. The text looks just as legible (even in the lower left) and the lower levels look exactly the same. When I drop the two images into a waveform monitor, they register the same in the shadow areas, not flatter or anything.The contrast looks exactly the same except that the highlights drop off slightly later on the ML one, which is exactly what we need - better highlights without making the image flatter.
--
How are the results stacking up Alex? Do we all have a similar preference?
Confirmed the issue in 720p. I think I'll just disable DIGIC
operations in these modes. It's a timing issue, but looks difficult to
solve.
--