10D high ISO revisited

I have been stating for a while now that the Canon 10D's ISO 1600 is
actually 800, under-exposed, and pushed a stop in the camera by doubling
the RAW data, and that Canon adds a pattern of horizontal lines that are
ones, zeros, or alternating every pixel for the least significant bit.
This much is true; however, I discovered a couple of things yesterday.

I set the 10D on a tripod, pointed it at my blinds which were closed and
illuminated from the outside, with a Sansevieria Cylindrica plant
sillouetted against it, and dark. I set the camera to ISO 3200, and set
exposure manually so that the blinds were at +3 stops. I took the shot
at 3200, and then dialed down the ISO as I took shots all the way down
to ISO 100, leaving the aperture and shutter speed fixed. I looked at
the RAW images from ISO 3200, 1600, and 800 in IRIS, using blackpoint
values from blackframes I also took at the same temperature, shutter
speed, and at all ISOs.

The RAW exposure was half at ISO 800 what it was at ISO 1600, as I
expected. This means that ISO 800 has a stop more highlight headroom, as
I've been saying for a while. HOWEVER, the ISO 1600 image had less
severe banding than the ISO 800 image, which means that the pattern of
odd and even least significant bits was done by Canon for a reason - to
reduce banding, being that the least significant bit was otherwise
unused. This means that, without a converter that eliminated banding, a
RAW file shot at ISO 1600 could be more immediately useful than an ISO
800 under-exposed and pushed, if you don't need that top stop of
highlights. This makes 1600 even more useful if you shoot JPEG than I
previously thought.

The other surprise was that the ISO 3200 RAW image was basically the
same as the ISO 1600 image. Same DR, same RAW values (ignoring random
noise differences) for the same absolute exposure. Apparently, the only
major difference between an ISO 1600 and an ISO 3200 RAW image on the
10D is the fact that there must be some kind of tag in the metadata that
tells a converter that ISO 3200 RAW files need to be scaled twice as
high in the converter.

I must have been fooled in the past by the fact that both ISO 3200 and
1600 RAW files from the 10D had only the least significant bit doctored;
I therefore assumed that ISO 3200 had ISO 1600 amplification, pushed a
stop. In fact, it seems to be ISO 800 pushed two stops, so ISO 3200 can
be considered totally useless, in general, to the RAW shooter, and 1600
is of marginal value because its RAW files are slightly de-banded
versions of 800 pushed to 1600 (of course, future RAW converters could
completely obsolete this value).

So, in summary, 1600 seems to be 800 pushed one stop in the camera with
de-banding "dither"; 3200 is the same, but with metadata that tells a
converter to push it a stop more.

The 20D has all zeros in its least significant bits, except for the
small number of ones that result from bad pixel interpolation.

Perhaps they dropped the proceedure because it took enough time to
affect burst speed in the 20D.
--

 

Straying from your post a bit, for the banding issue:

They could have simply used a random 1s or 0s for the LSB and the
banding would disapear.

Cheers,
Alan

 

In message <dpv1bp$m68$>,

Not really. The average of the line would be the same as it was
beforehand; all it would do is add a little more random noise; not
enough to hide the banding, and only a small percentage gain over the
existing random noise.
--