Luminous Landscape Forum
Equipment & Techniques => Cameras, Lenses and Shooting gear => Topic started by: ErikKaffehr on December 07, 2008, 03:51:16 pm
-
Hi,
I downloaded 3200 ISO raw images from http://www.imaging-resource.com/ (http://www.imaging-resource.com/) for both the Sony Alpha 900 and the Canon 5DII. I converted both to TIFF using "DCRAW" http://www.cybercom.net/~dcoffin/dcraw/ (http://www.cybercom.net/~dcoffin/dcraw/) .
I compared the TIFF files at actual pixels in LightRoom and they were in the same league, after that I profiled both images using Noise Ninja and found that the 5DII had a Noise Index of 69 (lum: 24, chroma: 45) whereas the Alpha 900 had a NI of 69 (Luma:26, chroma: 44) so I guess that RAW files are quite similar regarding noise, at least when processed by DCRAW.
Actual pixel view from Lightroom is enclosed, Canon left and Sony right. Visually I find that the Canon has a tighter noise pattern.
Erik
[attachment=10168:noise_3200.jpg]
-
Hi,
I downloaded 3200 ISO raw images from http://www.imaging-resource.com/ (http://www.imaging-resource.com/) for both the Sony Alpha 900 and the Canon 5DII. I converted both to TIFF using "DCRAW" http://www.cybercom.net/~dcoffin/dcraw/ (http://www.cybercom.net/~dcoffin/dcraw/) .
I compared the TIFF files at actual pixels in LightRoom and they were in the same league, after that I profiled both images using Noise Ninja and found that the 5DII had a Noise Index of 69 (lum: 24, chroma: 45) whereas the Alpha 900 had a NI of 69 (Luma:26, chroma: 44) so I guess that RAW files are quite similar regarding noise, at least when processed by DCRAW.
Actual pixel view from Lightroom is enclosed, Canon left and Sony right. Visually I find that the Canon has a tighter noise pattern.
Erik
[attachment=10168:noise_3200.jpg]
Interesting. I was expecting the 5DII to be a lot better than the Sony A900, given Sony are newer at this game. From the samples it looks like the difference is marginal.
Quentin
-
Actual pixel view from Lightroom is enclosed, Canon left and Sony right. Visually I find that the Canon has a tighter noise pattern.
[attachment=10168:noise_3200.jpg]
Erik,
Does 'tighter noise pattern' mean less noise?
Are you sure that both exposures we're equally exposed to the right? I think it's a bit inconclusive to base any opinion on a single comparison. The two images you compared were probably taken on different days and possibly by a different person. Whilst I'm sure Imaging Resource tries to be as consistent as possible, discrepancies and inaccuracies can sometimes take place.
There have been so many reports of the A900 being noisier than than the 1Ds3 and 5D2 above ISO 400, the only unresolved issue is perhaps how significant in practical terms is that noise difference. Perhaps not as significant as we imagine, especially when a converter such as Dcraw is used.
-
I downloaded 3200 ISO raw images from http://www.imaging-resource.com/ (http://www.imaging-resource.com/) for both the Sony Alpha 900 and the Canon 5DII. I converted both to TIFF using "DCRAW" http://www.cybercom.net/~dcoffin/dcraw/ (http://www.cybercom.net/~dcoffin/dcraw/)
I would be cautious with that, for
1. ACR does not handle the A900 files properly; I don't know how dcraw is doing that, nor do I know how exactly they should be treated.
2. The Sony A700 does carry out NR on the raw data; I don't think the A900 is different. In an earlier firmware version of the A700 this was not switchable; now it is, but I don't know if the file you are using was created with NR OFF.
3. Your Canon has a tighter noise pattern is quite an understatement; the A900 image is blotchy. This may come from the incorrect interpretation of the raw data, or from a botched noise reduction (already in-camera), or from the combination. It is unacceptable in my eyes, as it is now, but I am not sure if it can't be made much better with another raw processor. That time, when I was dealing with the A700's blotchiness, I was told that Sony's raw converter yields better result (but its workflow was not as good as ACR). It is possible, that the A900 image would look much better with Sony's converter. The ineptitude of some raw converters can not be held against a camera.
-
It is true that there were blotchy issues with high ISO A700 files when processed via Camera Raw. However, many users reported to us that this was resolved with a (relatively recent) firmware update provided by Sony. Gabor, I'm not sure if your experience with the A700 files was before or after the firmware update.
-
Here is the dark grey (not black) square of the color checker card from the ISO 3200 shots by the 5D2 and A900 (raw channels, not demosaiced). The exposures are almost identical in green; the spectral characteristics are very different, so the red and blue channels are not on the same level.
The "noise" of the A900 is horrendeous. I suspect, that some NR occured; thus the noise comparison between these shots is not valid.
-
It is true that there were blotchy issues with high ISO A700 files when processed via Camera Raw. However, many users reported to us that this was resolved with a (relatively recent) firmware update provided by Sony. Gabor, I'm not sure if your experience with the A700 files was before or after the firmware update.
I know about that issue. What I see with the A900 files is, that the result of the ACR conversion is still not linear. This is not a big issue on its own, except in very color sensitive cases, but the source of this problem - the incorrect black level compensation - contributes to the noise handling as well.
-
Hi,
Something I noted when downloading the files that the Sony file was just about 25 MBytes so it is obviously "compressed raw", because "full raw" would be about 37 MByte.
I think that DCRAW handles all files the same.
Best regards
Erik
I know about that issue. What I see with the A900 files is, that the result of the ACR conversion is still not linear. This is not a big issue on its own, except in very color sensitive cases, but the source of this problem - the incorrect black level compensation - contributes to the noise handling as well.
-
Ray,
I am not trying to prove anything. I'd just taken two images which should be pretty consistent and done some very simple checking. The reason I used DCRAW was simply that Lightroom 2.1 cannot open the Canon 5DII raw. DCRAW doesn't do any noise reduction and I think it essentially treats all raw files equal.
The reason that I wanted to checks this was that almost all results on the net are based on JPEG. I only use "raw". Also, according to DxO-mark there is not a lot of difference between sensors. Comparing the 1DsIII with the A900 I would say that the 1DsIII has a one step advantage, that is the 1DsIII would have the same noise as the A900 at the half ISO. I think that the A900 does noise reduction on "raw" and that this may be disabled in software, which probably was not done,
Finally I expected Panopeeper and some other knowledgeable guys to chime in and that obviously worked.
Erik
Erik,
Does 'tighter noise pattern' mean less noise?
Are you sure that both exposures we're equally exposed to the right? I think it's a bit inconclusive to base any opinion on a single comparison. The two images you compared were probably taken on different days and possibly by a different person. Whilst I'm sure Imaging Resource tries to be as consistent as possible, discrepancies and inaccuracies can sometimes take place.
There have been so many reports of the A900 being noisier than than the 1Ds3 and 5D2 above ISO 400, the only unresolved issue is perhaps how significant in practical terms is that noise difference. Perhaps not as significant as we imagine, especially when a converter such as Dcraw is used.
-
Something I noted when downloading the files that the Sony file was just about 25 MBytes so it is obviously "compressed raw", because "full raw" would be about 37 MByte.
The 25 MB is the lossily compressed size. However, do not connect this with the noise issue, as the lossiness sets in at pixel level 2000, but the noisy pixels (those, which did not get enough light) are usually well under that level.
-
Gabor,
What is your opinion only "lossily compressed" raw, better to be avoided or a good solution?
Erik
The 25 MB is the lossily compressed size. However, do not connect this with the noise issue, as the lossiness sets in at pixel level 2000, but the noisy pixels (those, which did not get enough light) are usually well under that level.
-
What is your opinion only "lossily compressed" raw, better to be avoided or a good solution?
You need to know, that the A900 generates only about 3950 levels, as opposed to 15000-16000 of Canon's and Nikon's 14bit depth. With the lossy compression, only 256 levels are kept, but not linearly; the way of lossiness resembles the mapping on an RGB space.
I personally find using such a high-end camera this way contradicting the principle of wanting to achieve the highest quality, but there are others with very different opinions.
-
Yes, but it was your opinion I was asking for. I essentially share the same view, but I have not found any discussion where the formats were compared.
Thanks!
Erik
You need to know, that the A900 generates only about 3950 levels, as opposed to 15000-16000 of Canon's and Nikon's 14bit depth. With the lossy compression, only 256 levels are kept, but not linearly; the way of lossiness resembles the mapping on an RGB space.
I personally find using such a high-end camera this way contradicting the principle of wanting to achieve the highest quality, but there are others with very different opinions.
-
You need to know, that the A900 generates only about 3950 levels, as opposed to 15000-16000 of Canon's and Nikon's 14bit depth. With the lossy compression, only 256 levels are kept, but not linearly; the way of lossiness resembles the mapping on an RGB space.
I personally find using such a high-end camera this way contradicting the principle of wanting to achieve the highest quality, but there are others with very different opinions.
Are we talking about D3 14 bits, or the D300/D3x 14 bits, which looks very much like it's just oversampling? 14bits is useless at this stage of the game.
-
Are we talking about D3 14 bits, or the D300/D3x 14 bits, which looks very much like it's just oversampling?
I don't know anything of the hardware, but I don't see any sign in the raw data indicating, that the 14bit data of the D300 is not "true".
14bits is useless at this stage of the game
Well, I disagree without being able to prove my opinion, which is, that the 14th bit is probably useless, but the 13th can be useful (perhaps the reason for 14 is the same as for why there is no row 13 on the airplanes and story 13 in many highrises :-).
Perhaps more sophisticated software will be able to utilize the extra levels, for example in noise reduction.
-
Are we talking about D3 14 bits, or the D300/D3x 14 bits, which looks very much like it's just oversampling? 14bits is useless at this stage of the game.
I've tested the numbers between D300 12 bits and 14 bits. As expected, there is a difference. You have naturally less noise with 14 bits and a bit more DR, which is probably insignificant with basic image processing. But, noise reduction on 14 bits apposed to 12 bits could make a big real world difference. Also, algorithms like shadows/highlights will show better toning on 14 bits. Anything the does extreme processing will benefit from 14 bits, even less noise after basic gamma.
-
Interesting. I was expecting the 5DII to be a lot better than the Sony A900, given Sony are newer at this game. From the samples it looks like the difference is marginal.
Quentin
They're not really new though. Remember the first 8MP camera--Sony F828?
-
The 25 MB is the lossily compressed size. However, do not connect this with the noise issue, as the lossiness sets in at pixel level 2000, but the noisy pixels (those, which did not get enough light) are usually well under that level.
What's odd here is that the A900 files are ALL about 25mb. The Canons vary a lot more, my 5D can be between 9 and 15mb depending on subject matter. When I tested the A900 a few months back I was surprised to see how they varied in size by only about 1mb must be a very different compression algorithm.
-
What's odd here is that the A900 files are ALL about 25mb. The Canons vary a lot more, my 5D can be between 9 and 15mb depending on subject matter. When I tested the A900 a few months back I was surprised to see how they varied in size by only about 1mb must be a very different compression algorithm.
My 1DS3 is variable too from 19 to almost 35MBs.
-
What's odd here is that the A900 files are ALL about 25mb. The Canons vary a lot more, my 5D can be between 9 and 15mb depending on subject matter. When I tested the A900 a few months back I was surprised to see how they varied in size by only about 1mb must be a very different compression algorithm.
The consistent raw file sizes has been true for the KM7D and Alpha 700 that I have, so makes sense A900 is the same. the compressed raw looked interesting, but I couldn't see the advantage when it lost bits to do it - storage is cheap these days :-)
-
Nick, all the Canon files are losslessly compressed with the result being that smooth images are small and high-variance images are big. At ISO 100, an image of tree bark (lots of fine detail) will be much bigger than an image of blue sky. For a fixed subject, the variance goes up with higher ISO, so the file size goes up too. i.e., in general higher ISO --> bigger files.
-
I've tested the numbers between D300 12 bits and 14 bits. As expected, there is a difference. You have naturally less noise with 14 bits and a bit more DR, which is probably insignificant with basic image processing. But, noise reduction on 14 bits apposed to 12 bits could make a big real world difference. Also, algorithms like shadows/highlights will show better toning on 14 bits. Anything the does extreme processing will benefit from 14 bits, even less noise after basic gamma.
Those results aren't in question. The issue is that it may not have anything to do with 14bits, but rather it's just an oversampling, and that's why it's cleaner(and why it takes longer to process.) Emil can weigh in here for a techical explanation of what I'm referring to
-
What's odd here is that the A900 files are ALL about 25mb. The Canons vary a lot more, my 5D can be between 9 and 15mb depending on subject matter. When I tested the A900 a few months back I was surprised to see how they varied in size by only about 1mb must be a very different compression algorithm.
Of course the file sizes are almost equal. The only difference in the file size comes from the embedded JPEG, which is around 500 KB for the A700, depending on the scenery and noisiness.
The A700/A900 lossy raw data is not compressed, thus its length is always 4288x2856 bytes for the A700 and 6080x4048 for the A900. The compression of the original raw data is achieved by just the lossiness. This is in contrast to the Nikons, which compress the lossy data as well (but the lossiness is much less with the D3 and D300 than with the A700 and A900).
-
Still, it isn't exactly a clear-cut case as I noticed that the Sony has less jaggies/pixelation on the edges of objects(as would be expected). Also, ISO3200 isn't something that I'd use more than once every few months. But I do desire more resolution more often, so there has to be a weighting towards the higher resolution image, IMO.
Yes, it would be better if it were 14 bit, but the colors are very good and when you come back down to ISO 800 or lower, it looks very nice.
http://www.cameralabs.com/reviews/Sony_Alpha_DSLR_A900/ (http://www.cameralabs.com/reviews/Sony_Alpha_DSLR_A900/)
This site has some examples of the NR on it - and you can see how great a difference it makes at each level. I actually prefer the NR on the highest setting myself(usually I don't - go figure).
-
Those results aren't in question. The issue is that it may not have anything to do with 14bits, but rather it's just an oversampling, and that's why it's cleaner(and why it takes longer to process.) Emil can weigh in here for a techical explanation of what I'm referring to
Well, I guess that by oversampling you mean that Nikon simply maps 12bit data from the ADC to 14bits. Right? Based on the fact the D300 drops frame rate when set on 14 bit to 2.5 FPS, I'd say it is true 14 bits out of the ADC.
-
Well, I guess that by oversampling you mean that Nikon simply maps 12bit data from the ADC to 14bits. Right? Based on the fact the D300 drops frame rate when set on 14 bit to 2.5 FPS, I'd say it is true 14 bits out of the ADC.
I don't understand what is meant. As I understand it, a film scanner can perform multiple readings of the same region (i.e. pass light through again and read the signal) and averaging these readings will reduce the random noise.
I doubt the D3x camera repeatedly re-exposes the sensor. So presumably the sensor is exposed, and charge builds up. Then maybe they try reading the charge in each sensor pit 4 times in order to halve the noise associated with the read path. So maybe a 12 bit read path but read multiple times to reduce noise. Who knows.