Sony A700 Dynamic Range test: 9.5 f-stops

[Guillermo Luijk]

Guillermo,
Is there anything wrong with the following approach? I stick a target on a white wall; test chart, newspaper, whatever. I then take a number of shots, halving the exposure each time.

I start off with say 1 sec at f8 and ISO 100 (or base ISO) and which is a full exposure to the right. After halving the exposure 10 times, I'm down to a 1,000th sec at f8. If I can still see some recognisable detail in the shot at a 1,000th sec (after processing in ACR) but not at 1/2,000th, then I can claim my camera has 10 stops of DR. Would this be correct?
[a href=\"index.php?act=findpost&pid=160153\"][{POST_SNAPBACK}][/a]

IMO would be perfectly correct. Better than my appreciations, as I depend on the assumption of sensor linearity in the shadows to correctly allocate elements in each EV zone. Your system would assure that you know the exact f-stop relation between your different shots.
I would use different targets however: test card, a tile with texture, a piece of cloth with texture and a piece of wood for instance, that would give more information to a photographer.

It's the same as I told Pano I am planning to do, but in my case with a plain card so that numeric SNR calculations can be done. But I am thinking it could be fine to put the white card together with some other textured stuff, so we have both the numeric calculation and the subjetive appreciation of noise.

[Ray]

Do you know of any specific example for that? I know of specific examles, which do not increase the dynamic range, and never heard of any others.
[a href=\"index.php?act=findpost&pid=160176\"][{POST_SNAPBACK}][/a]

I thought that was the idea behind Fuji's Super CCD employing two photodiodes at each site, one of which is less sensitive than the other. The smaller, less sensitive photodiode compresses the highlights by virtue of the fact it is less sensitive, not because of non-linear encoding.

The DR of the Fujifilm S5 Pro is described by dpreview as being about 11.8 stops, and that's in a jpeg.

[Panopeeper]

it would be fed with a pile of RAW shots over a uniform card at different exposures

This is not a generally accepted method, for you can not trust the consistency of the exposure (how reliable is your camera? You don't know it; my 20D is totally unreliable). Furthermore, even lighting is a true problem.

The accepted method is shooting a step wedge, i.e. you use a single shot and that encompassed a dozen or so stops, then you pick the extremes.

I don't find the link, but you can search for Stouffer wedge T4110.

[Guillermo Luijk]

I thought that was the idea behind Fuji's Super CCD employing two photodiodes at each site, one of which is less sensitive than the other. The smaller, less sensitive photodiode compresses the highlights by virtue of the fact it is less sensitive, not because of non-linear encoding.

The DR of the Fujifilm S5 Pro is described by dpreview as being about 11.8 stops, and that's in a jpeg.
[a href=\"index.php?act=findpost&pid=160186\"][{POST_SNAPBACK}][/a]

Yes, it is as you say. But both sensors are linear, so there is no compression in the RAW files, simply one of the sensors (the S) is 3.6 more sensitive than the other (the R). I am glad to read tjat figure since I studied in deep a RAW file from the S3 Pro and found that DR was, at least, 11 f-stops (I claim at least since the scene had a DR 11 f-stops, so there was nothing else to record on it).

The JPEG of the Fuji is another story, but the point is in the RAW structure.

[Panopeeper]

I thought that was the idea behind Fuji's Super CCD employing two photodiodes at each site, one of which is less sensitive than the other

Ok, I would not have described this is "highlight compression". I was thinking of some Nikon cameras, which apply a true highlight compression (a curve), when compressed raw data is requested - but that is occuring after the facts, i.e. it truely compresses the already captured image into less levels, but the same DR.

[Ray]

It's the same as I told Pano I am planning to do, but in my case with a plain card so that numeric SNR calculations can be done. But I am thinking it could be fine to put the white card together with some other textured stuff, so we have both the numeric calculation and the subjetive appreciation of noise.
[a href=\"index.php?act=findpost&pid=160184\"][{POST_SNAPBACK}][/a]

I'll give it a try with my 5D some time. I think an ideal target would be black text on white paper, but different sizes. At some point the small text will become unreadable, but the large text should remain legible until the lower limit of the DR is reached.

[Guillermo Luijk]

This is not a generally accepted method, for you can not trust the consistency of the exposure (how reliable is your camera? You don't know it; my 20D is totally unreliable). Furthermore, even lighting is a true problem.

The accepted method is shooting a step wedge, i.e. you use a single shot and that encompassed a dozen or so stops, then you pick the extremes.

I don't find the link, but you can search for Stouffer wedge T4110.
[a href=\"index.php?act=findpost&pid=160189\"][{POST_SNAPBACK}][/a]

I had the same thoughts as you (I had a feeling the time intervale of my camera's shutter could not very precise), but they are: this is a series of shots 1EV apart, and the log histograms calculated accordingly. As you can see the camera's shutter is as precise as a Swiss clock, it maaged to allocate each shot (non exposure corrected, non WB) in the same precise position of each f-stop of the linear sensor DR:






Anyway, in the way I am planning to calculate the SNR from the pile of RAWs, I don't need the exposure data (the EXIFF info I mean), I don't even need the shots to be exacly 1EV apart of to be uniformly apart, since I am going to calculate the exposure myself, by analysing the levels found in the RAW file. This assumes sensor linearity, which cannot be 100% true as we approach the deepest shadows.


I did not understand this: "The accepted method is shooting a step wedge, i.e. you use a single shot and that encompassed a dozen or so stops, then you pick the extremes.", could you explain a bit?

[Panopeeper]

As you can see the camera's shutter is as precise as a Swiss clock, it maaged to allocate each shot (non exposure corrected, non WB) in the same precise position of each f-stop of the linear sensor DR

I am really envious (though I have not tested this with the 40D yet), for this is sometimes important when shooting panoramas.

However, you want to compare cameras, so all of them need to be so accurate.

"The accepted method is shooting a step wedge, i.e. you use a single shot and that encompassed a dozen or so stops, then you pick the extremes.", could you explain a bit?

Bill Janes (who is posting on these forums as well) posted this on an Adobe forum:

Stouffer wedge

The steps are 1/3 stop apart.

[Jonathan Ratzlaff]

When I look at your original image, I don't think there is 10 stops of dynamic range there.  When you take a meter reading of the curtain, and then a meter reading of the shadow, is there 10 stops difference?  If not you need a subject with a higher range of light to dark,  say overcast sky and shadows in the room.  
Just wondering

[Guillermo Luijk]

I am really envious (though I have not tested this with the 40D yet), for this is sometimes important when shooting panoramas.

hehe my camera is a cheap 350D, so maybe yours is even better.

There is something interesting about those histograms I plotted, regarding sensor non-linearities near saturation: if you look closely to the histograms above, you can see some "distortion" in the shape of the blue channel histogram in the last f-stop (0EV) when compared to the others. The shape is slightly different to the others (it is a bit wider), but nothing very important actually.

Besides we must think that in that shot ALL R and G channels were blown, that means the sensor was working with 75% of the total pixels saturated and all blue pixels were surrounded by saturated pixels. I don't know the exact consequences of this regarding the electronic design of the sensor, but it seems reasonable to think that If under those extreme conditions the sensor was capable of yielding a resonable good quality in the B channel, we could conclude that non-linearities near saturation are negligible for regular photographic applications, where the last f-stop will be by far less occupied than in this example.

[Guillermo Luijk]

When I look at your original image, I don't think there is 10 stops of dynamic range there.  When you take a meter reading of the curtain, and then a meter reading of the shadow, is there 10 stops difference?  If not you need a subject with a higher range of light to dark,  say overcast sky and shadows in the room.  
Just wondering

Well it all depends on how linear the Sony sensor is. But assuming linearity, and this is quite reasonable at least for the first 8-9 f-stops, I assure you there were all those f-stops.

Think that I am showing you a gamma corrected image; this is not exactly would you would have seen (felt) if been there, but an approximation.

[douglasf13]

Well it all depends on how linear the Sony sensor is. But assuming linearity, and this is quite reasonable at least for the first 8-9 f-stops, I assure you there were all those f-stops.

Think that I am showing you a gamma corrected image; this is not exactly would you would have seen (felt) if been there, but an approximation.
[a href=\"index.php?act=findpost&pid=160313\"][{POST_SNAPBACK}][/a]

Hey, GLuijk.  I assume you used ISO 200 for the A700 right?  ISO 200 has the best dynamic range for this camera, and is it's true base ISO.

[Guillermo Luijk]

Hey, GLuijk.  I assume you used ISO 200 for the A700 right?  ISO 200 has the best dynamic range for this camera, and is it's true base ISO.

Right, it was ISO200 (although I did not shoot the test, just did the analysis).

[marcmccalmont]

sure, do it trough yousendit.com at gluijk(at)hotmail.com
Make sure you shoot a high dynamic range scene (like those with the bright window and very dark room). It's important that many textured (wood, clothes, floor tiles) objects in different degrees of shadow appear on the scene. It's also recommended if you use a tripod in order not to confuse noise with hand blur (what's the term for this in English?).
[{POST_SNAPBACK}][/a]

GLuijk, Hi I finally got home and took 3 shots of a scene with wide dynamic range.
G9, 5D and P30 here are the links to the RAW’s


G9
[a href=\"http://www.yousendit.com/transfer.php?action=download&ufid=735DA63E5B3E9730]http://www.yousendit.com/transfer.php?acti...35DA63E5B3E9730[/url]

5D
http://www.yousendit.com/transfer.php?acti...3758E200AAFF7F9

P30
http://www.yousendit.com/transfer.php?acti...9626CCD45C0359B

[John Sheehy]

Guillermo,
If the Sony A700 really does beat the 5D in terms of high ISO noise and dynamic range, and at least equals that of the D3, I'll definitely be getting one.
[a href=\"index.php?act=findpost&pid=160024\"][{POST_SNAPBACK}][/a]

Don't get too excited.  The A700's RAW data is far from RAW, at least at high ISOs.  It is filtered for noise, hence the low noise levels.  That means lower detail levels, especially in the deep shadows, which are nearly ironed flat in ISO 1600 "RAW" data.  I haven't looked at the low ISO shots, but they could be filtered too, but only in the deepest shadows.

Perhaps a new step needs to be taken in measuring noise, considering the increasing dishonesty of manufacturers; measuring whether or not it is filtered.  One approach that has been suggested is to bin the RAW data from flat, OOF areas and see if the noise diminishes at the rate at which it is supposed to, because if the noise is already filtered, it will not decrease as much as it should.  Of course, all banding components must be removed first because they are known not to reduce like 2-dimensional noise.

[Jonathan Wienke]

Perhaps a new step needs to be taken in measuring noise, considering the increasing dishonesty of manufacturers; measuring whether or not it is filtered.  One approach that has been suggested is to bin the RAW data from flat, OOF areas and see if the noise diminishes at the rate at which it is supposed to, because if the noise is already filtered, it will not decrease as much as it should.

Another advantage of my method of defining the noise floor (using my DR test chart framed so the center square is 100 pixels across) is that noise-reduction methods that smear detail will not result in an increase in measured DR, because the detail smearing will make the small text illegible just as easily as an increase in noise levels.

[John Sheehy]

Another advantage of my method of defining the noise floor (using my DR test chart framed so the center square is 100 pixels across) is that noise-reduction methods that smear detail will not result in an increase in measured DR, because the detail smearing will make the small text illegible just as easily as an increase in noise levels.
[a href=\"index.php?act=findpost&pid=162467\"][{POST_SNAPBACK}][/a]

OK, I'll agree to that (as a means to determine if filtering has ocurred, even if not to the rest of your reasoning for testing pixels as the definitive DR), but in a way, the *image* resolution is still the bottom line.  For example, with a 1.28GP camera, would you care if the noise was filtered, as much as you would with a 1.3MP camera?

[Panopeeper]

GLuijk, Hi I finally got home and took 3 shots of a scene with wide dynamic range.
G9, 5D and P30 here are the links to the RAW’s

Mark,

I downloaded the images and will analyze them. First I had to convert the P30 image in DNG format, for my program can not process the original raw file, and I found something strange.

Do you know, why a baseline exposure correction of -1EV is recorded? It is *not* exposure bias, at least that is zero in the original raw. Adobe's DNG converter inserted this correction value; it extracted that from the manufacturer specific data and I wonder why.

I saw other shots from P30, where it is +1 EV, so it is nothing universal with this camera.

[marcmccalmont]

Mark,

I downloaded the images and will analyze them. First I had to convert the P30 image in DNG format, for my program can not process the original raw file, and I found something strange.

Do you know, why a baseline exposure correction of -1EV is recorded? It is *not* exposure bias, at least that is zero in the original raw. Adobe's DNG converter inserted this correction value; it extracted that from the manufacturer specific data and I wonder why.

I saw other shots from P30, where it is +1 EV, so it is nothing universal with this camera.
[a href=\"index.php?act=findpost&pid=162605\"][{POST_SNAPBACK}][/a]

[marcmccalmont]

Mark,

I downloaded the images and will analyze them. First I had to convert the P30 image in DNG format, for my program can not process the original raw file, and I found something strange.

Do you know, why a baseline exposure correction of -1EV is recorded? It is *not* exposure bias, at least that is zero in the original raw. Adobe's DNG converter inserted this correction value; it extracted that from the manufacturer specific data and I wonder why.

I saw other shots from P30, where it is +1 EV, so it is nothing universal with this camera.
[a href=\"index.php?act=findpost&pid=162605\"][{POST_SNAPBACK}][/a]

The P30 is new to me and I found the Phase One labeling of a RAW file  ".tif" strange. I also found that because of this breeze browser doesn't handle the files correctly. Both ACR and C1 v4 seem to handle the conversion OK. I am liking the added dynamic range of the P30 seems shadows are much cleaner than the 5D but getting colors looking good is a bit of a challenge.
Marc

[attachment=4360:attachment]