Paul Gardner wrote:
For those who insist that it doesen't help, or don't remember the 8 bit processors (8080), go take a course in electrical engineering.
I can see where this will lead without any guidance (from an [electrical] engineer, perhaps? ). Pondria, do you have any suggestions as to how such samples should be shot in order to try to answer your question? It might help get things moving in the right direction more efficiently. We all probably have ideas in our head as to how such example shots might be captured, but discussing it beforehand might help.
Since getting back from my trip, I've had virtually no time to sit down yet with the 1200 CR2 files I have and start to evaluate IQ issues. I can say, however, that looking at the files, the main difference that is unquestionably obvious to me is that the darker tones in the images are - I don't know - lighter? I know that sounds sort of elementary, but while preserving highlights, the shadows look more true-to-life to me, like there is less contrast in the image, but sort of only in the shadow areas. Where the 1Ds files would have had deep shadows, the files from the MkIII simply look more natural to me, more like the actual scene that I photographed. I have no idea whether this is born of some fundamental change in the tone curve that they've implemented, but in general, the shots do look quite different than either 1Ds or 30D files.
Unfortunately, I only have a 30D for comparison, and really wish I still had the 1Ds for grins. I'd also really like to see a 1D MkIII entry for your tabulated DR data, if anyone has the time to provide you with the (properly captured) images. Perhaps someone who tested one of their initial cameras (in your list) now has a MkIII?
Paul Gardner wrote:
Breaking it down, you can see that the quality of the sensor is determined by how well the light hits each cell and how much splatter to adjacent cells. This is mechanical. Controlled by microlens, filters, quality of the cell. etc.
The accuracy of the output file is controlled by how well the resulting voltage is processed.
Based upon some of the results I'm seeing from the MkIII, I'm thinking that the underlined statement above may be fundamentally related to Canon's 'under the hood' tweaking of the MkIII, and may be responsible for much more than a simple increase in high-ISO performance.
I need to spend more time with the RAW files to convince myself that I'm not seeing things...
Jeff wrote:
..., the main difference that is unquestionably obvious to me is that the darker tones in the images are - I don't know - lighter? I know that sounds sort of elementary, but while preserving highlights, the shadows look more true-to-life to me, like there is less contrast in the image, but sort of only in the shadow areas. Where the 1Ds files would have had deep shadows, the files from the MkIII simply look more natural to me, more like the actual scene that I photographed. I have no idea whether this is born of some fundamental change in the tone curve that they've implemented, but in general, the shots do look quite different than either 1Ds or 30D files. ...Show more →
Actually, I got very similar sensation after I calibrated ACR for my 1Ds2. The gist of it came from the simple fact that Linear curve needs to be selected to match all 6 gray levels. The built-in tone curves of DPP and ACR are tuned to provide contrast or punch.
Not sure if this helps, but this was shot at ISO 200, 1/8000th at f/2.8; note the full scale from the DPP dialog box. No adjustments, just took a screen shot while viewing the CR2.
Jeff wrote:
Not sure if this helps, but this was shot at ISO 200, 1/8000th at f/2.8; note the full scale from the DPP dialog box. No adjustments, just took a screen shot while viewing the CR2.
The capture clearly shows the 14 distctive stops. Considering that no camera that we tested broke the 10 stop barrier, this is an amazing accomplishment. Actually TOO GOOD. If the RAW file can be made available to me, I would like to load it up in my ACR and feel the diff.
Jeff, are you sure that the HTP is not enabled ?
Pondria wrote:
The capture clearly shows the 14 distctive stops. Considering that no camera that we tested broke the 10 stop barrier, this is an amazing accomplishment. Actually TOO GOOD. If the RAW file can be made available to me, I would like to load it up in my ACR and feel the diff.
Jeff, are you sure that the HTP is not enabled ?
I can send you the RAW file if you e-mail me your address (use the link below).
I just looked back, and for the previous image, HTP was 'on'. For the following image, it was 'off':
Edited by Jeff on Jun 19, 2007 at 07:20 AM GMT (Reason: sorted out the 'this' and 'that' for clarity)
Pondria wrote:
The capture clearly shows the 14 distctive stops.
So based on this histogram you know what the light meter reading was for the highlight and shadow detail in the original scene? Are you sure that these -10 and +4 values correspond to actual doublings of light?
Never having spent much time with DPP (let alone v.3), I certainly have no idea what that range really means, nor how Canon has defined the different data ranges for different cameras. Maybe 'DavidP' (or Alan) can shed some light, as his 1D3/1D2/5D thread would suggest he may have spent some time on this issue.
Thanks for the images Jeff. The bottom line for me is that my 20D can't come close to matching the range shown in the HTP file and that's all that matters in the end. See life really is easier for the ignorant and uninformed.
Hrow wrote:
Thanks for the images Jeff. The bottom line for me is that my 20D can't come close to matching the range shown in the HTP file and that's all that matters in the end. See life really is easier for the ignorant and uninformed.
I demonstrated a week ago how the HTP can be simulated with pre-1Ds3 cameras.
Hrow wrote:
Thanks for the images Jeff. The bottom line for me is that my 20D can't come close to matching the range shown in the HTP file and that's all that matters in the end. See life really is easier for the ignorant and uninformed.
Pondria wrote:
I demonstrated a week ago how the HTP can be simulated with pre-1Ds3 cameras.
It's the range in either the HTP or non-HTP file that impresses me.
Pondria wrote:
Reviweing the thread, multiple times it's been mentioned that 14bit has more head-room in in-camera or post processing.
I'm not convinced that this is the case. Embedded processors or RAW covnerters have all minimum 16/32 bit data width. 12 or 14 bit data need to be padded to 16 bit or 32bit anyway.
You may well be right that it is not the case, but it should be the case. There are several ways to fit 12 bits of raw image data into a 16 bit data format. The simplest and probably the most used in the past has been to add the extra bits at the most significant end (high end) of the possible data values and set those bits to zero so that they have no practical effect.
What I believe has happened in the case of the 1D3 is that all data values have been increased relative to their 12-bit sensor counterparts and then the meaning of those increased values has been decreased. This allows for extra tonal gradation without making the image any different in effective brightness. Obviously to do this some compensation has to be made in the camera and any software that trries to interpret the raw data or else the image may look too bright. In data terms this could be implemented by putting two zero bits at the lesat significant end of the data (causing a multiplication by two, twice) to form a 14-bit output from the sensor, and then putting just two zero bits at the most significant end. It may instead be done with a single eztra bit at the dark end so that another bit is left available for HTP (high tone priority) mode.
My contention is that ADC chips are so cheap (under $2.00 each in single quanties) that the cost of 16 bits is trivial. Why is Canon not just implementing them! I know they are trying to balance the battery life with the processing, BUT I'm sure I'm not alone in being able to scrafice some battery life for the improvement in IQ, especialy in the 1DS and 5D series. I'll settle for 200 Images per charge for landscapes. Let the PJs have their high speed 1000 shot cameras in the 1D series. Maybe Canon will come to their senses in the 1DSIII. I'm trying not to go to the MF cameras but the ZD is starting to look good, and I'm hoping that the ZDII will go from 14 to 16 bits.
Alan,
I'll use your terminlogy to get my thought across better to you.
I have been maintaing the position that if you just shift the 12bit data to the right by two bits to make it a 14bit data and pad the two LSBs with zero, it would be effectively the same as the 14bit data that 1D3 produces UNLESS 1D3 has improved S/N. Because the 2 LSBs that 14bit has WITHOUT the improved S/N would be just random numbers (noise) containing no relevant information for the image.
I've been asking for the data to show us if there is any S/N improvements. Jeff is going to send the raw files.
Paul Gardner wrote:
My contention is that ADC chips are so cheap (under $2.00 each in single quanties) that the cost of 16 bits is trivial. Why is Canon not just implementing them! .
I bet the ADC itself has been 16bit.
But the pipe line throughput ( measured in bits/second ) requirement chops off the unnecessary bits. No reason to crunch the padded-zeros and random numbers (noise).
Jeff wrote:
Never having spent much time with DPP (let alone v.3), I certainly have no idea what that range really means, nor how Canon has defined the different data ranges for different cameras. Maybe 'DavidP' (or Alan) can shed some light, as his 1D3/1D2/5D thread would suggest he may have spent some time on this issue.
Edited by Jeff on Jun 19, 2007 at 08:18 AM GMT
To my way of thinking the range is simply the range of raw image data brightness values, but don't confuse "raw image data" with "raw data". Raw image data begins with the raw data first presented by the image sensor (which includes the photon sensor and the built-in amplifiers and ADC) and then manipulated by the demosaicing algorithm needed to take individual single-colour values at each location and produce effective 12- or 14-bit RGB data values at each pixel location and then an effective brightness value at each location.
The sensor is a multi-stage device. It effectively receives light at each pixel and produces a corresponding analogue voltage signal representing the amount of light. That voltage is then amplified and then fed into an analogue to digital converter. The 1D3 ADC can produce a 14-bit output instead of 12 bits. The numbers represent the amount of light received through a coloured red, green or blue filter at each pixel - with just one colour per pixel location.
After the sensor produces this raw data the firmware and Digic III computer converts it raw image data, representing the RGB colour and brightness at each pixel location. This is demosaicing. The output from this is what is represented in the raw image histogram window. It is the basis for conversion to 8-bit jpeg rgb data or, outside the camera, to 8- or 16-bit TIFF rgb data.
The histogram shows the frequency of occurrence (vertical scale) vs data value (horizontal scale). In DPP 3 the middle tonal value - whatever number it happens to be in terms of brightness value (varies with camera model according to an arbitrary choice by Canon) - is assigned a value of zero and all other values are either positive (for relatively brighter) or negative (for relatively darker). The horizontal scale is limited at the high and low ends by how much brighter or darker the values can be, which depends on the data bit level of the sensor (12 or 14 bits).
For most DSLRs the bright limit is about +3.7 stops brighter than middle tone, give or take a little bit. The dark end limit for these cameras is 12 stops darker than the bright end limit. The 1D3 is pretty much the same in normal mode, as shown to me recently by DavidP and again by Jeff's screen dump images posted in this thread. (these were the first time I have been able to see this info as I do not have a 1D3). What is interesting is that the 1D3 in HTP mode extends both the non-HTP bright limit by 1 stop and the non-HTP dark limit by 1 stop, but the shape and position of the raw image histogram relative to the middle tone level is pretty much unchanged. Obviously DPP is already doing some interpreting of this data or it would show the histogram 1 stop darker.
What is also interesting is that the HTP tone conversion curve shown in the raw image histogram, which is the curve that maps the raw image data to the 8-bit RGB data values of a jpeg file, remains pretty much the same below about +3 stops, and is the same but tilted over a bit more (stretched) above that, when compared with the non-HTP curve. So it seems that no use is made of the extra data captured below about -5 stops, the point at which the tone curve bottoms out, in either HTP or non-HTP mode. That extra data can only be useful if the raw image data exposure level or contrast are altered
Where did the extra stop of bright data in HTP mode come from ? Assuming that the sensor already presents all available data values in normal mode up to the maximum possible 16,383 for 14 bits, there is no way it can find an extra stop of data at the bright end for the HTP mode. The limit of the ADC has already been reached. So HTP has to be fudged and is probably done partly in the sensor hardware rather than just in the firmware. They cannot just divide everything by two because that would decrease the maximum possible value too. i.e. nothing from the ADC would exceed 8,191. So I expect the analog output (noise plus data) of the sensor is only amplified half as much before the usual conversion to 14-bit digital is applied. Then the data all represents half of the measured light value and has to be doubled by the raw conversion software or firmware. If the software is not aware of HTP then the result will probably be a dark image with bright highlights.
My conclusions drawn from all of this are:
the 1D3 14-bit raw image data presents two more bits of significant data than 12-bit cameras.
In non-HTP mode the 1D3 raw data file is likely to present more tonal gradations in any exposure range, probably by a factor of 4 (2 extra data bits), than a 12-bit camera but it is only likely to be noticeable in the darker areas. This does not imply any extra, useful dynamic range; just more values to describe the same old range with consequently less posterisation.
In HTP mode the definition of middle tone and most other values are moved lower (darker) to be 1 stop further away from the maximum value that the 14-bit sensor is able to present. To compensate for this the data is then interpreted as being one stop brighter than the raw value indicated. After correct interpretation this leaves one extra stop of dynamic range at the bright end of the scale. This is relative to the non-HTP mode. There is still one extra bit of data (twice as many possible levels) to describe every brightness range compared with 12-bit cameras.
HTP mode offers up to one extra stop of dynamic range at the bright end (only) by using one extra data bit at the bright end, but offers half as many tonal gradations as non-HTP mode in any brightness range (still twice as many as 12-bit cameras).
Due to the tone conversion curve little or none of the data below about -5 stops can be used unless the raw data exposure level is increased prior to conversion.
). Pondria, do you have any suggestions as to how such samples should be shot in order to try to answer your question? It might help get things moving in the right direction more efficiently. We all probably have ideas in our head as to how such example shots might be captured, but discussing it beforehand might help.
