On Wed, Jul 4, 2012 at 9:08 AM, Stefan Stavrev <[email protected]> wrote:
> Kevin, can you please provide more details?
>
> You mean 10 bits per channel? That means 1024 intensity values, and width of
> histogram that much?
>
> And why would you want to search for missing bins? As far as I know the
> histogram is used to give you a bird eye view of the values distribution.

1 word - quality.

When working with images that under go lots of potential steps of
processing being read and written to disk between each step, small
errors can accumulate. Traditionally we have striven to provide 10
bits of integer 'log' coded (or rec 601/709 for video) from one end to
the other of our pipeline. if you have images that measure up to this,
then suddenly you have something that say operates at 8 bits, you will
notice a tooth comb like effect in the histogram, due to the loss of
the 2 bits. If instead you lost say bit 7 of the data because your A/D
converter on your scanner/camera was broken, you may not see it
initially, but after processing and colour correction you can start to
see odd effects like image banding show up. This will show up as a
difference spacing of missing bins, but still obvious in a histogram
if drawn correctly - typically I'd display bins 1:1 or even 2:1 (2X
bigger) if I have the screen space.

We have people here (and worse some of our clients :-) who can and
have spotted a single 10 bit digit difference in scanned film images,
when projected (reasonably noisy conditions which should 'mask' the
visibility of such things according to accepted dogma), especially if
it occurs in a structured way e.g. one half of an image, or one shot
in a sequence, or that section where something was roto'd by a 3rd
party vendor (all real examples :-). You can't always identify quite
what is wrong but you 'feel' a difference from personal experience.

These days as camera noise floors have dropped 10 bits precision is
not really enough - all digitally projected masters should really pass
a 12 bit per channel quality image check - whilst I don't think that
is currently possible due to a range of miss-matched inputs, outputs
and processing steps this will likely be needed in a few years.

Yes it is possible to automate the detection of missing codes and from
the pattern identify what is wrong, but artists and clients always
like pictures even if they are charts.

I don't disagree that a rough approximation view of a histogram to
show you the distribution 'shape' and position is also frequently
needed and can show things such as clipped images - abrupt ends to the
filled bins, often accompanied with sudden increases in counts i.e. a
spike.

Spikes in the middle of a histogram can show areas of flat colour
which are unnatural and can also stand out to somebody else after we
have finished on images. Showing a reduced bin count in this situation
is akin to filtering the spike away using a poor quality frequency
response like a 2D box filter does in image down sampling. This kind
of thing occurs when some texture/lighting  is missing from a render
resulting in solid colour areas which will tend to bias the
distribution towards some colours - histograms of images converted to
alternate colour spaces can be useful for spotting this sometimes.

Kevin
_______________________________________________
Oiio-dev mailing list
[email protected]
http://lists.openimageio.org/listinfo.cgi/oiio-dev-openimageio.org

Reply via email to