Ad
Ad
Ad
Pages: « 1 [2] 3 4 ... 7 »   Bottom of Page
Print
Author Topic: DR, DxO, DSLR, MFDB, CMOS, CCD  (Read 17856 times)
ErikKaffehr
Sr. Member
****
Offline Offline

Posts: 7680


WWW
« Reply #20 on: July 09, 2010, 12:45:38 PM »
ReplyReply

Feppe,

Emil is scientist and I am an engineer ;-)

Best regards
Erik

Quote from: feppe
The problem is you are arguing about what amounts to religion with science, which is bound to fail
Logged

fredjeang
Guest
« Reply #21 on: July 09, 2010, 12:46:55 PM »
ReplyReply

But then, comes immediatly a question.

If the D3x is that impressive, and in many aspects supposely matches or equal the highest MF backs,
considering its much lower price, its better distribution and universality, and that can even shoot video,
why the MF users are not selling their gears like hot-dogs and go all for the D3x?

(not talking about the one who use regularly view and tech cameras but MF cameras)

Logged
ErikKaffehr
Sr. Member
****
Offline Offline

Posts: 7680


WWW
« Reply #22 on: July 09, 2010, 12:59:21 PM »
ReplyReply

Hi,

The idea with this discussion is not compare D3X to P65+ or any other combination, but to find out if there is a reasonable explanation to the perceived 4-6 stop benefit in DR of MFDBs over DSLRs.

Best regards
Erik


Quote from: fredjeang
But then, comes immediatly a question.

If the D3x is that impressive, and in many aspects supposely matches or equal the highest MF backs,
considering its much lower price, its better distribution and universality, and that can even shoot video,
why the MF users are not selling their gears like hot-dogs and go all for the D3x?

(not talking about the one who use regularly view and tech cameras but MF cameras)
Logged

joofa
Sr. Member
****
Offline Offline

Posts: 488



« Reply #23 on: July 09, 2010, 01:28:23 PM »
ReplyReply

Quote from: Ray
Is there anyone who would claim that a camera which could not provide legible text at a specific underexposure, has a higher DR than another camera which can produce legible text in the same circumstances, using lenses of comparable quality?

It would be interesting to compare results using Jonathan Wienke's method.

IMHO, the above has little to do with DR. In engineering literature DR is normally understood to be a "point property", as opposed to a "neighbor hood property". For e.g., color is a point property - you have a well defined notion of the color at a given pixel. Texture is a neighborhood property - can you have texture at a single pixel? DR has generally been understood in terms of the dynamic range of a given pixel (point property). Attempts to extrapolate the meaning of DR from point property to neighborhood property have typically resulted in confusion. Some cases may still have merit, for example the (neighborhood property) notion of "image DR/SNR", and more importantly, "image scale dependent DR/SNR". However, the problem is that such notions are not constructed properly, with flawed assumptions, such as that erroneous notion that SNR for a natural image varies as square root of pixel summed/averaged, which adds further to the confusion.

Jonathan Weinke, or for that matter anybody else, is free to define a notion of "image quality", but relating that measure of IQ to DR, such as the above claim of "text-based DR" is not what has traditionally been understood in engineering circles.

Joofa
« Last Edit: July 09, 2010, 01:34:59 PM by joofa » Logged

Joofa
http://www.djjoofa.com
Download Photoshop and After Effects plugins
joofa
Sr. Member
****
Offline Offline

Posts: 488



« Reply #24 on: July 09, 2010, 01:44:12 PM »
ReplyReply

Removed.
« Last Edit: July 09, 2010, 01:56:21 PM by joofa » Logged

Joofa
http://www.djjoofa.com
Download Photoshop and After Effects plugins
fredjeang
Guest
« Reply #25 on: July 09, 2010, 02:42:44 PM »
ReplyReply

Quote from: ErikKaffehr
Hi,

The idea with this discussion is not compare D3X to P65+ or any other combination, but to find out if there is a reasonable explanation to the perceived 4-6 stop benefit in DR of MFDBs over DSLRs.

Best regards
Erik
Correct, I was way off topic with this one.  

Cheers.
Logged
ErikKaffehr
Sr. Member
****
Offline Offline

Posts: 7680


WWW
« Reply #26 on: July 09, 2010, 03:29:29 PM »
ReplyReply

Hi,

I don't argue with your view. On the other hand I'm not necessarily only interested in a technically accepted definition of DR but also perception of DR. Essentially in anything "for real" that can matter to a trained observer and that can be explained by theory or proven by samples.

Also, I think that Mark Dubovoy was talking about DR on details with full texture.

Best regards
Erik

Quote from: joofa
IMHO, the above has little to do with DR. In engineering literature DR is normally understood to be a "point property", as opposed to a "neighbor hood property". For e.g., color is a point property - you have a well defined notion of the color at a given pixel. Texture is a neighborhood property - can you have texture at a single pixel? DR has generally been understood in terms of the dynamic range of a given pixel (point property). Attempts to extrapolate the meaning of DR from point property to neighborhood property have typically resulted in confusion. Some cases may still have merit, for example the (neighborhood property) notion of "image DR/SNR", and more importantly, "image scale dependent DR/SNR". However, the problem is that such notions are not constructed properly, with flawed assumptions, such as that erroneous notion that SNR for a natural image varies as square root of pixel summed/averaged, which adds further to the confusion.

Jonathan Weinke, or for that matter anybody else, is free to define a notion of "image quality", but relating that measure of IQ to DR, such as the above claim of "text-based DR" is not what has traditionally been understood in engineering circles.

Joofa
Logged

LKaven
Sr. Member
****
Offline Offline

Posts: 832


« Reply #27 on: July 09, 2010, 04:09:18 PM »
ReplyReply

Quote from: fredjeang
That's exactly what I think is the problem.
How do we stritcly have to use a term.

I do not have the answer to your question, but it seems that DR embrasse something too large and therefore confuse.
I do not think that "dynamic range" is a singular entity, and I'd compare it more to terms like "intelligence" for which there is no singular measure, or a word like "game" (thank you Prof. Wittgenstein) in which there is no singularly necessary quality (or jointly sufficient quality) that any thing has to have in order to be a 'game'.  In actuality, it is practical characteristics of the data and its use that give us the impression of quality.  Sometimes we cluster a number of our impressions under the term "dynamic range" though in the end, it is our ability to make practical use that serves our interests.

If a file is malleable, and able to sustain significant amount of adjustment without falling apart, that is a practical benefit.  In the case of the D3x, the extreme low noise in the blacks gives the files an important practical benefit in all kinds of adjustments.  This is quite different from the painstaking care that needs to be taken with, say, 5DII files, which show all kinds of strange noise in the shadows, which must be hidden or otherwise mitigated in practical use.  If what the D3x has is "dynamic range" it's the kind I'm most glad to have.
« Last Edit: July 09, 2010, 04:10:17 PM by LKaven » Logged

ErikKaffehr
Sr. Member
****
Offline Offline

Posts: 7680


WWW
« Reply #28 on: July 09, 2010, 04:34:02 PM »
ReplyReply

Hi,

Just two comments:

1) Nikon D3X has much better DR than Canon 5DII according to DxO, this may be consistent with what you see

2) The DR measured by DxO doesn't say about the character of the noise just about the noise present

Best regards
Erik


Quote from: LKaven
I do not think that "dynamic range" is a singular entity, and I'd compare it more to terms like "intelligence" for which there is no singular measure, or a word like "game" (thank you Prof. Wittgenstein) in which there is no singularly necessary quality (or jointly sufficient quality) that any thing has to have in order to be a 'game'.  In actuality, it is practical characteristics of the data and its use that give us the impression of quality.  Sometimes we cluster a number of our impressions under the term "dynamic range" though in the end, it is our ability to make practical use that serves our interests.

If a file is malleable, and able to sustain significant amount of adjustment without falling apart, that is a practical benefit.  In the case of the D3x, the extreme low noise in the blacks gives the files an important practical benefit in all kinds of adjustments.  This is quite different from the painstaking care that needs to be taken with, say, 5DII files, which show all kinds of strange noise in the shadows, which must be hidden or otherwise mitigated in practical use.  If what the D3x has is "dynamic range" it's the kind I'm most glad to have.
Logged

bjanes
Sr. Member
****
Offline Offline

Posts: 2825



« Reply #29 on: July 09, 2010, 04:36:39 PM »
ReplyReply

Quote from: ejmartin
DxO does have the requisite information if you know where to look:
http://luminous-landscape.com/forum/index....c=42158&hl=

It is a bit exasperating to continually hear this complaint about DxO.  The points about noise character are well taken however.  They can be quantified by a more refined analysis of the noise spectrum, but DxO does not do this (at least, they don't present any results on their website).
As Emil has indicated in earlier posts, the engineering definition of DR places an upper limit on the DR defined by other criteria. The read noise and full well capacity (which determines shot noise) can be incorporated into a noise model that describes fairly well the observed noise characteristics of a camera, as Roger Clark has demonstrated on his web site.

The KODAK KAF-50100 is used in the Hasselblad H3DII-50, and its read noise is 12.5 electrons at the specified read out rate and the full well capacity is 40,300 electrons. The Nikon D3 (for which I have data) has a full well of 65568 electrons and a read noise at the base ISO of 17.63 electrons. One can plug these values into a spreadsheet using Clark's model and get the shown values. One can read the DR from the table as the exposure in f/stops from saturation for a given S:N. For a S:N of 1, the DR of the KAF-50100 is about 11.75 stops, which is in excellent agreement with the DXO measurement of 11.35 stops. The DR of the D3 at S:N = 1 is 12 stops as compared to the DXO value of 11.92 stops. The red values show where DR is limited by read noise.

[attachment=23050:SNR_DR.gif]

[attachment=23051:DXO_DR.GIF]

This is, of course, per pixel DR. The KAF-50100  has many more pixels with a smaller full well than that of the D3. The KAF-50100 collects 2.6 x more electrons than the D3, as can be determined by multiplying the full well times the pixel count. With the Hasselblad, one could print a larger picture at a given print resolution as compared to the Nikon, which after all is one of the main reasons for a MFDB over a dSLR. Alternatively, one could downsize to the pixel count of the Hasselblad to that of the Nikon and gain dynamic range by pixel binning. This is the rationale of the DXO print DR.

[attachment=23052:DXO_DR_print.gif

Although the D3 was a breakthrough for Nikon, it's DR at base ISO is not as great as it could be, since the read noise at low ISO is relatively high, but improves with higher ISO. As Emil has explained, this accounts for the flattening of the DR curve at low ISOs and is due to limitations in the electronics downstream to the sensor. This limitation is not present for the Hasselblad (or the Nikon D3x).

This discussion does not take pattern noise or the noise spectrum into account, but I would venture to hypothesize that a 4-6 stop improvement in DR of a MFDB as compared to the Nikon D3x or other top dSLRs is unlikely. The experts must be observing something other than DR when comparing the two formats. It would be interesting for them to demonstrate their claims with data or actual pictures. Until then, I remain skeptical.
« Last Edit: July 09, 2010, 04:42:42 PM by bjanes » Logged
joofa
Sr. Member
****
Offline Offline

Posts: 488



« Reply #30 on: July 09, 2010, 05:10:54 PM »
ReplyReply

Quote from: bjanes
As Emil has indicated in earlier posts, the engineering definition of DR places an upper limit on the DR defined by other criteria. .... This is, of course, per pixel DR. .... Alternatively, one could downsize to the pixel count of the Hasselblad to that of the Nikon and gain dynamic range by pixel binning. This is the rationale of the DXO print DR.

Do you know how DXO is calculating print DR?

An issue with the "downsizing to the pixel count"  problem is that, as I mentioned before, many authors, and it would appear Emil Martinec included (from the link that you cite in your message), have taken the notion of DR as in the case of a flat field and tried to extrapolate that on a natural image, thinking that the "scaling of noise" is valid in that case also, which it is not IMHO. Because, they appear to calculate the effect of such downsizing on noise statistics only, and not realizing that the signal is also getting affected (blurred), and both S and N in the ratio S/N are varying as opposed to just N as they seem to assume.

A more extensive model of DR/SNR variation in downsampling is needed and it is not too difficult to work out such models.

Joofa
Logged

Joofa
http://www.djjoofa.com
Download Photoshop and After Effects plugins
ErikKaffehr
Sr. Member
****
Offline Offline

Posts: 7680


WWW
« Reply #31 on: July 09, 2010, 05:22:43 PM »
ReplyReply

Hello Bill,

So you also say that the two sensors are about one stop apart, and this is what would be expected. Also that this is in agreement with DxO data. I'm not really surprised.

Best regards
Erik



Quote from: bjanes
As Emil has indicated in earlier posts, the engineering definition of DR places an upper limit on the DR defined by other criteria. The read noise and full well capacity (which determines shot noise) can be incorporated into a noise model that describes fairly well the observed noise characteristics of a camera, as Roger Clark has demonstrated on his web site.

The KODAK KAF-50100 is used in the Hasselblad H3DII-50, and its read noise is 12.5 electrons at the specified read out rate and the full well capacity is 40,300 electrons. The Nikon D3 (for which I have data) has a full well of 65568 electrons and a read noise at the base ISO of 17.63 electrons. One can plug these values into a spreadsheet using Clark's model and get the shown values. One can read the DR from the table as the exposure in f/stops from saturation for a given S:N. For a S:N of 1, the DR of the KAF-50100 is about 11.75 stops, which is in excellent agreement with the DXO measurement of 11.35 stops. The DR of the D3 at S:N = 1 is 12 stops as compared to the DXO value of 11.92 stops. The red values show where DR is limited by read noise.

[attachment=23050:SNR_DR.gif]

[attachment=23051:DXO_DR.GIF]

This is, of course, per pixel DR. The KAF-50100  has many more pixels with a smaller full well than that of the D3. The KAF-50100 collects 2.6 x more electrons than the D3, as can be determined by multiplying the full well times the pixel count. With the Hasselblad, one could print a larger picture at a given print resolution as compared to the Nikon, which after all is one of the main reasons for a MFDB over a dSLR. Alternatively, one could downsize to the pixel count of the Hasselblad to that of the Nikon and gain dynamic range by pixel binning. This is the rationale of the DXO print DR.

[attachment=23052:DXO_DR_print.gif

Although the D3 was a breakthrough for Nikon, it's DR at base ISO is not as great as it could be, since the read noise at low ISO is relatively high, but improves with higher ISO. As Emil has explained, this accounts for the flattening of the DR curve at low ISOs and is due to limitations in the electronics downstream to the sensor. This limitation is not present for the Hasselblad (or the Nikon D3x).

This discussion does not take pattern noise or the noise spectrum into account, but I would venture to hypothesize that a 4-6 stop improvement in DR of a MFDB as compared to the Nikon D3x or other top dSLRs is unlikely. The experts must be observing something other than DR when comparing the two formats. It would be interesting for them to demonstrate their claims with data or actual pictures. Until then, I remain skeptical.
Logged

BernardLanguillier
Sr. Member
****
Offline Offline

Posts: 8211



WWW
« Reply #32 on: July 09, 2010, 06:07:17 PM »
ReplyReply

Quote from: fredjeang
That's exactly what I think is the problem.
How do we stritcly have to use a term.

I do not have the answer to your question, but it seems that DR embrasse something too large and therefore confuse.
Remember that horsepowers have also the same issue: there are 2 measurements.
The Data horsepower and the real horsepower. They are not the same.
This has been used in car engineering to define 2 concepts that where shocking together.

Fred,

Hum... I am afraid that won't cut it.

If Mark tells us he sees 6 stops gap in DR what we have is a clearly measurable entity, otherwise why write "6"? So there must be a clear definition.

Is "6" a metaphor for "a lot"? Fine, but then how does "a lot" translate in practical terms? What is the characteristic of these files showing "a lot" more DR?

Cheers,
Bernard
Logged

A few images online here!
BernardLanguillier
Sr. Member
****
Offline Offline

Posts: 8211



WWW
« Reply #33 on: July 09, 2010, 06:10:29 PM »
ReplyReply

Quote from: ejmartin
DxO does have the requisite information if you know where to look:
http://luminous-landscape.com/forum/index....c=42158&hl=

It is a bit exasperating to continually hear this complaint about DxO.  The points about noise character are well taken however.  They can be quantified by a more refined analysis of the noise spectrum, but DxO does not do this (at least, they don't present any results on their website).

Emil,

I am not questioning DxO. I am just trying to understand what some of the folks here mean when they say that the DxO figure doesn't reflect the actual real world DR they see in their files.

Until now the only concrete aspect that has been mentioned is the look of noise. So in essence does "the 6 stops more DR" just point to the fact that the 5DII has banding issues in deep shadows?

Cheers,
Bernard
Logged

A few images online here!
BernardLanguillier
Sr. Member
****
Offline Offline

Posts: 8211



WWW
« Reply #34 on: July 09, 2010, 06:18:29 PM »
ReplyReply

Quote from: fredjeang
But then, comes immediatly a question.

If the D3x is that impressive, and in many aspects supposely matches or equal the highest MF backs,
considering its much lower price, its better distribution and universality, and that can even shoot video,
why the MF users are not selling their gears like hot-dogs and go all for the D3x?

(not talking about the one who use regularly view and tech cameras but MF cameras)

Off topic indeed, but so much fun.

If the backs are so much better, why are like new second hand 39 megapixel backs selling at the same price point as the D3x?

More seriously, there are many valid reasons why people might want to keep using backs regardless of their DR:
- ADDED: they have signficantly more resolution (I had stupidely omited the obvious )
- They have the platform and lenses and like them,
- They like the look of AA filter less files,
- They are used to subtle DoF effects related to the large size of the sensor,
- Backs can be mounted on a variety of plarforms (including LF cameras),
- The sensor is easier to clean,
- Backs feel/look more profesional,
- It creates a differentiator relative to other photographers unable to afford a back,
- They like thethered shooting for which backs have been optimized,
- They have a personnal relationship with their back dealer that they would be unable to find with Canon/Nikon,
- They are able to meet Phaseone R&D guys and to influence their directions but wouldn't be able to do the same with Canon/Nikon,
- ...

That is what I find most irritating about this whole DR story. Backs don't need to oversell their DR using mythical and un-provable qualities, they have enough going for themselves already.

Cheers,
Bernard
« Last Edit: July 11, 2010, 06:16:54 PM by BernardLanguillier » Logged

A few images online here!
BartvanderWolf
Sr. Member
****
Offline Offline

Posts: 3771


« Reply #35 on: July 09, 2010, 06:38:16 PM »
ReplyReply

Quote from: joofa
Do you know how DXO is calculating print DR?

http://www.dxomark.com/index.php/eng/Our-publications/DxOMark-Insights/Detailed-computation-of-DxOMark-Sensor-normalization

Cheers,
Bart
Logged
joofa
Sr. Member
****
Offline Offline

Posts: 488



« Reply #36 on: July 09, 2010, 07:22:04 PM »
ReplyReply

Quote from: BartvanderWolf

Hi Bart,

Thanks for the link. IMHO, it appears that DXO is making the same mistake that I have pointed out in my earlier message, i.e., just to consider noise statistics in image size reduction. For a flat field it will work. However, for a natural, real, image, that is not correct as I see it, since the signal is also affected in the reduction operation and as I said before both S and N in S/N ratio are changing. Please note that SNR will typically improve in downsizing, however, the noise only statistics only sets the upper bound on that number as it corresponds to a flat field. For a natural image, one can't ignore the change in signal in addition to noise in such SNR calculations.

Joofa
« Last Edit: July 09, 2010, 07:24:52 PM by joofa » Logged

Joofa
http://www.djjoofa.com
Download Photoshop and After Effects plugins
EricV
Full Member
***
Offline Offline

Posts: 132


« Reply #37 on: July 09, 2010, 08:16:53 PM »
ReplyReply

Anyone seriously interested in pursuing this topic should first read and understand the excellent post by ejmartin Dynamic Range and DxO.  Continuing from there ...

Noise at extremely low signal levels is likely to be dominated by sensor electronics readout.
[blockquote]Suppose you are interested in "Engineering DR", defined by (maximum signal) / (signal equivalent to dark noise).  If your camera electronics has pixel readout noise of 40 electrons and each pixel has full well (saturation) capacity of 40000 electrons, then your DR is (40000/40) = 1000 = 60db = 10 f/stops.  The camera manufacturer can improve this DR for you either by lowering electronics noise or by increasing pixel full well capacity. [/blockquote]
Noise at moderate signal levels is likely to be dominated by photon statistics.  
[blockquote]Suppose you are interested in "Photographer DR", defined by S/N>16 in the darkest pixels of the image which still contain acceptable detail.  Then you require those dark pixels to contain at least 256 electrons.  If the pixels have maximum full well capacity of 40000 electrons, then your DR is (40000/256) = 156 = 44db = 7.3 f/stops.  The camera manufacturer can improve this DR for you by increasing the pixel full well capacity.  Increasing the pixel count will also improve DR, assuming you are interested in "print DR" rather than "pixel DR". [/blockquote]
On a slight tangent to this topic, DR can also be improved by merging multiple exposures, at least for stationary subjects and a tripod mounted camera.  I think some clever camera manufacturer should provide a one-button press function which captures and merges two exposures taken with different sensor integration times.  This could easily beat the DR of any existing camera.
« Last Edit: July 09, 2010, 08:36:28 PM by EricV » Logged
ejmartin
Sr. Member
****
Offline Offline

Posts: 575


« Reply #38 on: July 09, 2010, 11:05:22 PM »
ReplyReply

Quote from: ErikKaffehr
Emil,

What is your point on my assumption that a significant advantage (like 4-6 stop) in DR for MFDBs would also be seen as good high ISO performance?

Yes you have that correct, and it's a very good point to make.
Logged

emil
ejmartin
Sr. Member
****
Offline Offline

Posts: 575


« Reply #39 on: July 09, 2010, 11:11:45 PM »
ReplyReply

Quote from: joofa
IMHO, the above has little to do with DR. In engineering literature DR is normally understood to be a "point property", as opposed to a "neighbor hood property". For e.g., color is a point property - you have a well defined notion of the color at a given pixel. Texture is a neighborhood property - can you have texture at a single pixel? DR has generally been understood in terms of the dynamic range of a given pixel (point property). Attempts to extrapolate the meaning of DR from point property to neighborhood property have typically resulted in confusion. Some cases may still have merit, for example the (neighborhood property) notion of "image DR/SNR", and more importantly, "image scale dependent DR/SNR". However, the problem is that such notions are not constructed properly, with flawed assumptions, such as that erroneous notion that SNR for a natural image varies as square root of pixel summed/averaged, which adds further to the confusion.

Jonathan Weinke, or for that matter anybody else, is free to define a notion of "image quality", but relating that measure of IQ to DR, such as the above claim of "text-based DR" is not what has traditionally been understood in engineering circles.

Joofa

On the contrary, it has a lot to do with DR.  Engineering DR is quite close to the range of tonal values over which pixel SNR exceeds unity, which condition is closely tied to the ability to extract details at the pixel scale.
Logged

emil
Pages: « 1 [2] 3 4 ... 7 »   Top of Page
Print
Jump to:  

Ad
Ad
Ad