O-MD E-M5 - The RawDigger Base ISO Challenge (Continued)

Started Apr 19, 2012 | Discussions thread
Flat view
Anders W Forum Pro • Posts: 21,466
O-MD E-M5 - The RawDigger Base ISO Challenge (Continued)

Since the old thread with the same title "maxed out", I take the liberty of starting a new one in order to answer some questions asked of me in the last few posts.

Detail Man wrote:

Anders W wrote:

The GH2 file has now arrived and I have had a chance to look at it. As I suspected, the noise distribution is badly clipped. About 7/8 of the cases are in the first (leftmost) bin.

So, in order to compute an estimate for the GH2 directly corresponding to that we already have for the E-M5, I suggest we take recourse to the data available from Sensorgen (aka bobn2), which are in turn computed on the basis of the DxOMark measurements. According to Sensorgen, the max level (saturation point) is 11803 and the read noise 6.0, both expressed in electrons rather than ADUs but that makes no difference as long as the same unit is used for both quantities. So, according to the engineering definition, the GH2 has a DR of

2log(11803/6) = 10.94

as against 11.90 for the E-M5. Consequently, we are left with the 1 EV (or so) difference between the two cameras that we have already talked about above.

Well, I think that I (maybe) understand what you are getting at about the GH2's distribution. But when the EM5 distribution is tightly-placed at low values, we are impressed ... but not (similarly) in the case of the GH2 ? What do you mean by "badly clipped" ? I am open to substantive evidence.

Perhaps I shouldn't have talked just about clipping here. What I mean is that if the distribution had looked like just like the right half of a normal distribution, we would have had a decent basis for assuming that the underlying distribution is normal with the clipping point at the center. We might then have applied the correction factor of 1.66 appropriate for estimating the standard deviation of a normal distribution based on one of its halves only. But in a variety of ways (see previous posts), it doesn't look like the right half of a normal distribution. In view of that, we do not, in my opinion have much ground for assumptions about what the left half actually looks like or for deriving correction factors based on those assumptions.

Note: This thread will be maxed, so perhaps we should indeed start a continued thread to further converse a bit.

As you can see, I just did.

How do you know it to be "clipped" (implying that it has not been internally truncated, or "cut off", at the center of the noise-distribution) ?

I cannot know that for sure. I can only guess. The problem is that I have no grounds for assuming that it has indeed been cut at the center of the underlying distribution or that the left half looks just like a mirror image of the right half. Even if I made those assumptions, it would be tricky to come up with a correction factor that fits the assumed distribution since it has a rather peculiar (rather than standard) shape.

It is certainly the case that if the EM5's (SD or RMS, etc.) statistic of 1.0 is divided by the 1.66 factor that you have presented for such distributions, then (I calculate) the difference between the EM5 and the GH2 statistics result in being a mere 0.3 EV (or so) different (with the EM5 being that much lower-noise than the GH2). This is certainly less flattering than the range of 1.0 EV that your method is yielding - but is that a fair approach to take (in the zero-illumination case, anyway) ?

No, not in my opinion. We have no reasonable basis for making the assumptions on which we could base our use of the 1.66 correction factor. Moreover, better data (those measured by DxOMark and later used by Sensorgen to calculate the key statistics we need) are available than those we get from a black-frame shot with the GH2.

Do understand your reasoning about the more general analysis, but wonder whether the GH2 stats ought to be summarily discarded based only on the fact that most of it's area is in the 1st bin

See above.

We have no way of knowing if Panasonic has failed to place one-half of the distribution above zero - and when the EM5 distribution looks "tight" (populated near the zero-point), we are impressed. Is this a fair way to compare the (zero light illumination) data that we have for the two cameras ?

The major reason why I have confidence in using the E-M5 data for DR calculations is not that it is tight but that its general shape conforms very well to what one would generally expect the error distribution to look like. The major reason why I don't feel confident using the GH2 distribution for the same purpose has already been spelled out above.

 Anders W's gear list:Anders W's gear list
Panasonic Lumix DMC-G1 Olympus OM-D E-M5 Olympus E-M1 Panasonic Lumix G Vario 14-45mm F3.5-5.6 ASPH OIS Panasonic Lumix G Vario 7-14mm F4 ASPH +28 more
Panasonic Lumix DMC-GH2
If you believe there are incorrect tags, please send us this post using our feedback form.
Flat view
Post (hide subjects) Posted by
Keyboard shortcuts:
FForum PPrevious NNext WNext unread UUpvote SSubscribe RReply QQuote BBookmark MMy threads
Color scheme? Blue / Yellow