DP1 Review on DCR Compares Leica M8 to DP1 on page 20

[AlanG]

Advertisement (gone after registration)

Alan, I think I agree with you - a B&W test chart should cover all options. Basically your are testing all colors at once.

 

The fuzzy images in the test for me are close to being incredible, are they really sure there was no focus error? The output quality of the M8 would be inferior if this test was realistic (and we know M8 isn't inferior in image quality certainly not at the lower ISO settings). Did they use DNG or jpg?

 

I don't see how there could be a focus error as all of the color charts are shot at once. So how could one be blurry and the rest sharp?

 

It was reported here that they used jpegs. If so, I think that could be the problem.

[Photoskeptic]

I believe the real problem with the tests is that they are using JPEG. Most of us on this forum already know how weak the M8 is when shooting jpeg's and don't

[mjh]

I believe the real problem with the tests is that they are using JPEG. Most of us on this forum already know how weak the M8 is when shooting jpeg's and don't

JPEG usually (but not neceessarily) applies chroma-subsampling which reduces the spatial resolution of colour. Still, the DP1 with its Foveon sensor uses JPEG as well and performs noticably better than Bayer-sensor cameras when there is colour contrast, but little luminance contrast. Moreover, even the cameras using a Bayer pattern still perform quite well as long as there is still some contrast in the green channel. I don’t see how JPEG compression could explain the difference in performance (in fact, using JPEG only puts the DP1 at a disadvantage, as JPEG’s chroma-subsampling is detrimental to the Foveon sensor’s improved spatial resolution in the colour channels).

[mjh]

I don't see how there could be a focus error as all of the color charts are shot at once. So how could one be blurry and the rest sharp?

In theory it could, but only if the colour pigments used did reflect a lot of IR while the black and white of the b&w chart reflected next to none. Not too likely in my view, even assuming that the amount of IR could suffice to explain the blur, which I doubt. And how likely is it that IR reflectivity should affect those colours the most that happen to be under-represented in a Bayer pattern, namely red and blue?

[jaapv]

I don’t think IR has got anything to do with it. As has already been observed, this would affected the b&w chart as well, but in fact, this is the sharpest of all the six charts, even in the red and blue channels.

 

Presumably, this is just an effect of the demosaicing algorithm used. Given that only 25 percent of all the pixels are sensitive to red (or blue), one should expect the red channel to be blurry. However, any reasonably sophisticated demosaicing algorithm will use the data from the higher resolution green channel to improve the sharpness in the other two channels. Whenever there is a change in the amount of red from one red-sensitive pixel to the next, and a change in the amount of green at roughly the same spot, the algorithm will assume that the gradient for the red channel is just as steep as that for the green channel, even when the red-sensitive pixels are spaced wider apart. Obviously, this works best for b&w patterns where changes in the amount of red, green, and blue are perfectly correlated. It still works when there is sufficient contrast in the green channel that corresponds with the patterns in the red and/or blue channels, but when even the green channel fails to provide the necessary detail, the M8’s demosaicing algorithm apparently resorts to a conservative approach of merely interpolating the missing values in the red and blue channels. In fact, the differences in sharpness observed in the red and blue channels are just a more extreme version of what we see in the green channel: good contrast in the green channel enables the demosaicing algorithm to make the red and blue channel resolution to look better than it actually is, but when even the green channel fails, we see what resolution for red and blue is really like with a Bayer sensor.

 

 

Apparently, the Nikon DSLRs used for comparison in the review apply a less conservative algorithm. Looking at the red/blue chart, it seems obvious that neither the D60 nor the D40x offer better resolution, only the patterns show more contrast and appear to be sharper. This is merely what you get from sharpening. In fact, the low-contrast pattern recorded by the M8 looks cleaner that the results obtained with the Nikon models. It has to be said, though, that the sharpened look of the Nikon images is quite typical of what you get from modern DSLRs; it’s not just Nikon. Leica’s conservative approach to demosaicing is more unusual.

 

That is very clear, Michael. But what, if any, application has this test got to real-life results then?

[mjh]

But what, if any, application has this test got to real-life results then?

Not that much, to be fair. It illustrates nicely what Foveon sensors are good at, and sensors with a Bayer pattern less so. One can also deduce some properties of the demosaicing algorithm from the results, but that is largely academic. Personally, I happen to like the scientific approach, isolating the contributing factors such as the primary colours on the one hand and luminance on the other. But it would take some explaining to point out what the relevance of these effects would be.

[barjohn]

Advertisement (gone after registration)

I wonder if it is because the demosaicing algorithm is optimized around a high contrast black and white image and that results in less than optimum on the two color combinations.

[mjh]

I wonder if it is because the demosaicing algorithm is optimized around a high contrast black and white image and that results in less than optimum on the two color combinations.

It is the Bayer pattern that is optimized for black and white patterns, or patterns with a lot of luminance contrast anyway. The demosaicing algorithm has to struggle with those cases the Bayer pattern was not optimized for, such as red/blue contrast.

[blakley]

I'd be interested in repeating the test myself under various conditions, Alan. Do you know where I could find a copy of the color resolution target DPReview uses?

[jaapv]

Not that much, to be fair. It illustrates nicely what Foveon sensors are good at, and sensors with a Bayer pattern less so. One can also deduce some properties of the demosaicing algorithm from the results, but that is largely academic. Personally, I happen to like the scientific approach, isolating the contributing factors such as the primary colours on the one hand and luminance on the other. But it would take some explaining to point out what the relevance of these effects would be.

 

So it probably would make more sense to publicize such a test in a more specialized environment than a general digiphoto website. But then the rigour of the methodology would need to be more firmly anchored.

[AlanG]

In theory it could, but only if the colour pigments used did reflect a lot of IR while the black and white of the b&w chart reflected next to none. Not too likely in my view, even assuming that the amount of IR could suffice to explain the blur, which I doubt. And how likely is it that IR reflectivity should affect those colours the most that happen to be under-represented in a Bayer pattern, namely red and blue?

 

DPReview stated this about the lighting."The IR filter simply would have made NO difference in this test. We used our normal daylight simulation softboxes which use high CRI bulbs at about 5000 Kelvin." So that seems like a fluorescent light source which would put out a minimal amount of IR if any.

 

Considering that the light reflecting from the target is much great than any amount of IR radiation that could possibly be reflecting from it, I don't see how it could ever constitute such a blurry image.

 

I think a lot of IR radiation in an otherwise "normal" scene could at most make a slight halo around a sharp image but that's it. Otherwise the M8 would work pretty terribly without filters in many situations.

 

As for getting your own colored test charts, you should contact DPReview. I don't know who makes those charts.

 

But you could probably use your imagination and design some sort of simple test if your goal is only to see if the IR filter makes any difference with red/blue edges on jpegs and raw.

[barjohn]

I have searched for these test charts and the only reference I can find for where to get them is from FOVEON. I asked Phil for the source but he is now ignoring me, apparently because this is being discussed on a foreign forum (from his). Here is a link to another test using the 5D versus the SD 14 (same sensor).

 

Sigma SD14 Resolution: Can it Hang with the Big Dogs?