Ad
Ad
Ad
Pages: « 1 ... 6 7 [8]   Bottom of Page
Print
Author Topic: What chance has Sigma's SD1?  (Read 30724 times)
joofa
Sr. Member
****
Offline Offline

Posts: 488



« Reply #140 on: June 03, 2011, 09:26:57 AM »
ReplyReply

Example, a challenging task in computer vision is object recognition, say just identifying if a particular picture has a cat in it.

Joofa  

How could I use a lowly word, such as an "object", in relation to a cat? My sincere appologies to anyone whose feline sensibilities are hurt.

Joofa
Logged

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

Posts: 966



« Reply #141 on: June 03, 2011, 09:39:44 AM »
ReplyReply

If there is a significant diffrrence between a Foveon and Bayer of similar pixel counts, the real culprit is not the Bayer CFA, but the anti-alias filter.

easy to test - take any MF(DB) image (no AA) and zoom 100% to a pixel 1:1 level, then do the same w/ Foveon image (not SD1 - take production camera like DP2 for example) and have the subjects of the same size (in pixel terms - subject that occupies the same amount of pixels on both sensors in the image) of course... at that pixel peeping level MF(DB) is destroyed by a simple P&S DP camera, regardless of the absence of AA... there is no replacement for displacement (demosaicking vs no demosaicking).... printing or resizing (down to 4.6mp of the production Foveon) of course will show the different picture, there the sheer amount of sensels that MF(DB) has will give it an advantage, but not @ 100% zoom on screen
« Last Edit: June 03, 2011, 09:44:45 AM by deejjjaaaa » Logged
BartvanderWolf
Sr. Member
****
Offline Offline

Posts: 3767


« Reply #142 on: June 03, 2011, 10:22:15 AM »
ReplyReply

easy to test - take any MF(DB) image (no AA) and zoom 100% to a pixel 1:1 level, then do the same w/ Foveon image (not SD1 - take production camera like DP2 for example) and have the subjects of the same size (in pixel terms - subject that occupies the same amount of pixels on both sensors in the image) of course... at that pixel peeping level MF(DB) is destroyed by a simple P&S DP camera, regardless of the absence of AA... there is no replacement for displacement (demosaicking vs no demosaicking).... printing or resizing (down to 4.6mp of the production Foveon) of course will show the different picture, there the sheer amount of sensels that MF(DB) has will give it an advantage, but not @ 100% zoom on screen

Apart from the differences in optics, that seems to be a fair test. Have you done it? Is that how you reached your conclusion of "at that pixel peeping level MF(DB) is destroyed by a simple P&S DP camera"?

All I've seen produced by others are assumptions. I've personally done a test of the effect that demosaicing has on Luminance resolution, and it's not that dramatic as some would like us to believe. Only in the case of equal luminance will a loss of Luminance resolution be seen, because modern demosaicing algorithms tend to optimize more for Luminance resolution than for Chroma resolution.

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

Posts: 488



« Reply #143 on: June 03, 2011, 10:53:24 AM »
ReplyReply

I've personally done a test of the effect that demosaicing has on Luminance resolution, and it's not that dramatic as some would like us to believe.

Here is a little different actual test of "resolution" with and without demosaicing on real images on two exact same sensors, with the only difference being the presence of CFA on one, and the other being mono.



AA filter: None.
IR filter: None.
Optics: Same paramters between the two; such as f-stop, exp time, etc.
Demosaic method: Bilinear interpolation.

Note that bilinear interpolation is not the best method, hence the two images kind of show two extremes with a CFA and without.

Joofa
« Last Edit: June 03, 2011, 10:55:36 AM by joofa » Logged

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

Posts: 3767


« Reply #144 on: June 03, 2011, 11:58:07 AM »
ReplyReply

Note that bilinear interpolation is not the best method, hence the two images kind of show two extremes with a CFA and without.

Hi 'Joofa',

Indeed, clearly visible but not all that earth-shattering especially with a bit of sharpening added to your JPEG.

Cheers,
Bart
Logged
Aku Ankka
Newbie
*
Offline Offline

Posts: 26


« Reply #145 on: June 03, 2011, 12:01:11 PM »
ReplyReply

Here is a little different actual test of "resolution" with and without demosaicing on real images on two exact same sensors, with the only difference being the presence of CFA on one, and the other being mono.

Demosaic method: Bilinear interpolation.

Note that bilinear interpolation is not the best method, hence the two images kind of show two extremes with a CFA and without.

Hardly a test of CFA effect on image quality, but a good example on what a very simple interpolation algorithm achieves from the recorded data.
« Last Edit: June 03, 2011, 12:10:31 PM by Aku Ankka » Logged
ejmartin
Sr. Member
****
Offline Offline

Posts: 575


« Reply #146 on: June 03, 2011, 12:15:01 PM »
ReplyReply

Bilinear will soften edges a lot more than a typical demosaic method used in any good raw converter, as well as introduce a lot more aliasing.
Logged

emil
Aku Ankka
Newbie
*
Offline Offline

Posts: 26


« Reply #147 on: June 03, 2011, 12:17:18 PM »
ReplyReply

For what it is worth, here are a a couple of sample images attached, taken from different distances to allow for comparison of details. Also included a 200% vs. 500% comparison.

Camera used: Pentax K20D - so AA-filter and Bayer CFA  were present during imaging Wink
Target was a 20 euro bill - blue-red areas to make resolving more difficult for the demosaicing algorighm. Raw-conversion was done with Lightroom 3 with all settings zeroed, ie. no sharpening what so ever.

Edit: added a sharpened version of the small one just to make it clearer how well details are resolved.

To my eyes it seems like a sensor with a Bayer CFA will resolve quite nicely even when unfavorable detail color.
« Last Edit: June 04, 2011, 05:53:16 AM by Aku Ankka » Logged
joofa
Sr. Member
****
Offline Offline

Posts: 488



« Reply #148 on: June 03, 2011, 01:20:37 PM »
ReplyReply

Hardly a test of CFA effect on image quality ...

It is not a test of how good the CFA response can be made using different demosaic algorithms, if that is what you mean here. The intent was to show the kind of two "extremes", i.e., any good demosaicing algorithm will be operating within these two loose end points.

Joofa
Logged

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

Posts: 488



« Reply #149 on: June 03, 2011, 01:29:10 PM »
ReplyReply

Hi 'Joofa',

Indeed, clearly visible but not all that earth-shattering especially with a bit of sharpening added to your JPEG.

Cheers,
Bart


Hi "Bart",

True indeed. There is an interesting evaluation of different deMosaic algorithms on the following link:

http://www.libraw.org/articles/bayer-moire.html

The author recommends LMMSE for further analysis given its good response on "closer to reality" images (author's words). And, which is not surprising given the blurring effect of AA + other effects on an image before demosaic happens as drawn in a simplified linear fashion below:



Joofa
Logged

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

Posts: 575


« Reply #150 on: June 03, 2011, 07:51:20 PM »
ReplyReply

The typical test of demosaic algorithms is to take a three-color image (eg from a scan), 'forget' 2/3 of the colors leaving only those in the Bayer pattern, and then run the resulting mosaic image through the demosaic algorithm of choice.  The original image provides the reference 'true' image, the mosaic/demosaic route provides a test of the faithfulness of the demosaic image to that 'truth'.  In principle one can also test the effects of OLPF filtering by applying the relevant anti-alias blur operation between mosaic and demosaic (though I have not done that on the images below).

Here is a standard test image often used in the literature on demosaic algorithms:



and here is what comes from a good demosaic algorithm applied to the mosaiced image one gets by forgetting all but the RGGB pattern from each quartet of pixels:

Logged

emil
NikoJorj
Sr. Member
****
Offline Offline

Posts: 1063


WWW
« Reply #151 on: June 04, 2011, 03:57:26 PM »
ReplyReply

I'd say that it would be more useful to have a sensor that allows to output something that resembles what the human eye saw, with minimal processing.
I'd say at least that with our limited chroma resolution, we generally may not remark that an image has a chroma resolution a bit behind its luminance resolution?
« Last Edit: June 04, 2011, 03:59:39 PM by NikoJorj » Logged

Nicolas from Grenoble
A small gallery
BartvanderWolf
Sr. Member
****
Offline Offline

Posts: 3767


« Reply #152 on: June 04, 2011, 05:12:44 PM »
ReplyReply

Here is a standard test image often used in the literature on demosaic algorithms:

[SNIPPED images for brevity]

and here is what comes from a good demosaic algorithm applied to the mosaiced image one gets by forgetting all but the RGGB pattern from each quartet of pixels:

Hi Emil,

Very nice demonstration, thanks. Is that with your full version of AMaZE in MatLab?
My old test was done with Aqua, but it's not around for download anymore.

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

Posts: 575


« Reply #153 on: June 04, 2011, 05:24:49 PM »
ReplyReply

Hi Emil,

Very nice demonstration, thanks. Is that with your full version of AMaZE in MatLab?
My old test was done with Aqua, but it's not around for download anymore.

Cheers,
Bart

Yes.  Actually, in Mathematica rather than Matlab.  As one can see, it does decently well with high frequency, greyscale detail such as the picket fence.  Errors are largest with the colored edges of the life preserver and the yellow rope since the details are largely in more sparsely sampled reds; and in the grass (more random high frequency structure is harder to infer than regular, more patterned structure).
Logged

emil
BartvanderWolf
Sr. Member
****
Offline Offline

Posts: 3767


« Reply #154 on: June 04, 2011, 07:10:49 PM »
ReplyReply

I'd say at least that with our limited chroma resolution, we generally may not remark that an image has a chroma resolution a bit behind its luminance resolution?

Hi Nicolas,

While true, things aren't as black and white (pun intended).

Restricted to the Fovea centralis, which is occupied mostly with cones (for colorvision), human visual color resolution peaks when enough light is available. Color information by itself usually has a low spatial frequency contribution to our vision and in scenes of natural objects (it's more of a warning signal at a distance, because e.g. Red may be warm/hot/blood). That is easily demonstrated in a Lab (or HSB or HSL) respresentation of an average image, there is little resolution to be seen.

However, the human visual system (HVS) depends heavily on pattern detection (to avoid information overload) for detail discrimination, and therefore luminosity as a means to convey the important (edge detail) message is rated higher in our brain. So counting cones and/or rods alone is not telling the whole story. Contrast, or MTF, is very important for our detail resolving abilities.

The HVS, which involves the brain, therefore has a variable resolution depending on average scene brightness and luminosity contrast. The brilliant weighting of a Bayer CFA is a very good (although not perfect) instrument to allow the most relevant information to be recorded. The fact that the cones of our eyes are more randomly distributed complicates a direct comparison with an aliasing prone ordered sampling method.

Nevertheless, as also Emil has demonstrated above, it is possible to achieve a visually very good/convincing reconstruction of the actual scene with only a faction of the data (which saves a lot of storage and speeds up the storade (frames/second) and transfer of data). In fact, when one analyses the R, G, and B, MTF resolution of an image (e.g. in Imatest) with adequate luminance contrast, the RGB resolutions of a Bayer CFA image are virtually identical to the luminance resolution. Only in the worst case scenario of chroma resolution in the absence of most luminosity contrast info will there be a benefit for full RGB resolution/sampling. When enlarging the resulting image, there is an obvious, although limited, benefit to having as much accurate (also RGB) data available.

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

Posts: 8908


« Reply #155 on: June 05, 2011, 02:27:44 AM »
ReplyReply

Of course, it could be that the real reason why Sigma priced the DSP1 at US$9,700 was to generate some pre-publicity.

This thread is an example of such pre-publicity. The camera hasn't even been released, but we have 150 odd posts on LL discussing its possible properties.
Logged
bill t.
Sr. Member
****
Offline Offline

Posts: 2693


WWW
« Reply #156 on: June 06, 2011, 05:32:55 PM »
ReplyReply

Take heart, you can pre-order from from BhPhoto for an amazingly affordable $6,899!  Seems like a bargain.  When did you say the 5d MKIII is coming out?
Logged
Pages: « 1 ... 6 7 [8]   Top of Page
Print
Jump to:  

Ad
Ad
Ad