The Megapixel Myth

[KBerntsen]

Here's an interesting blog I got of Ken Rockwell's site.

http://pogue.blogs.nytimes.com/2006/...ogues-posts-2/

[Bartman01]

One issue I can see with that 'test':

A 5mp image created by a 5mp camera will have more/larger noise/defects than a 5mp image downressed from a 13mp image. Downressing 'cleans up' noise/defects as it tosses out fine detail, so you are printing a 'cleaner' image.

[TXhummer2]

That is exactly why Real Life sometimes looks better in photos....

[KBerntsen]

Originally Posted by Bartman01

One issue I can see with that 'test':

A 5mp image created by a 5mp camera will have more/larger noise/defects than a 5mp image downressed from a 13mp image. Downressing 'cleans up' noise/defects as it tosses out fine detail, so you are printing a 'cleaner' image.

I was having the same thought. I would have liked it better if it had been different images. With different cameras.

[Wes]

Interesting, but there are flaws in the test. Those have been pointed out here and on the site.

[jeffkohn]

On the show, we did a test. We blew up a photograph to 16 x 24 inches at a professional photo lab. One print had 13-megapixel resolution; one had 8; the third had 5. Same exact photo, down-rezzed twice, all three printed at the same poster size. I wanted to hang them all on a wall in Times Square and challenge passersby to see if they could tell the difference.

This test is fundamentally flawed. They took a 13mp image and down-rezzed it to 5mp, which is hardly the same thing as taking a 5mp image with a 5mp camera.

Add to that we know nothing about the camera/lens used in this test, shooting methodology, etc. I mean sure, if you take a D2x and stick a $50 Quantaray lens on it and proceed to shoot with sloppy hand-holding technique, the results aren't likely to showcase what the D2x is capable of. Image selection also will have a big impact here, as some images have much more fine detail and would make differences in resolution more apparent.

I don't disagree that too much is made of megapixels in product marketing, especially with regard to small-sensor point-n-shoots where the pixel density is getting ridculous and signal-to-noise ratios are getting worse and worse. But I can guarantee you that there's a difference in final print quality between my D70 shots and my D2x shots even at 12x18, let alone 16x24.

[venchka]

Conversely, last night I saw with my own two eyes a HUGE (40x30 maybe) enlargement of a 5 MP image from a high end Olympus fixed lens DSLR. The image was printed here in Houston by Thomas Reprographics, the NW Freeway location. I'm assuming they have some fancy, high end, pixel adding software. From the proper viewing distance the image was outstanding. I was impressed.

[Tapper]

To understand this, lets do a thought experiment.

Lets say you have 3 pictures of exactly the same scene, with exactly the same colors, same luminosity, same aperature, same lens resolving power, and same viewport:

The largest image raster will contain the most detail. The smallest raster will contain the least detail. But - the dimensions of objects and areas of color will be identically the same. That is to say - a line 10mm long will be proportionally 10mm long in each photo. But in the largest raster, the edges of that line will have more contiguous detail - because the larger raster has more pixels available to register the detail. Importantly, the line will be located in the same position in each raster, proportional to the size of the raster.

Here's the thing - if you reduce the resolution using the same algorithm for each image, then the images will converge at each resolution. That is to say, if you reduced a 3500x3000 pixel raster to 2500x2142, then the images will be almost identical assuming an algorithm with perfect edge detection. Any differences, will be due to mathematical straightening of edges in the formerly larger picture. Thus, reducing the 3500x3000 pixel image to 2500x2142 results in two almost identical images, but the image that starts out at 2500x2142 will not have been smoothed by interpolation - and thus will be the better image (it has not thrown off detail yet).

As you reduce again, lets say to 1500x1285, the middle resolution image is now smoothing, and rapidly converging into an image which is identical to the originally largest image. That is to say - as the edges are smoothed, they approach a point at which they must occupy the exact same pixels in order to represent the physical dimensions of the objects or areas in the original image.

Eventually, there will be a convergence point at which all three images will be identically and exactly the same, no matter what resolution they were shot at.

Now obviously, it isn't possible to take three images at exactly the same light, absolute position, etc. But the smoothing effect still happens, and the effects on detail hold true.

So the big question is - where does the printer raster land in this continuum? If the printer raster is too small, then all the images will have converged to the point that the amount of detail is pretty much the same. If the raster is large, then the largest images will produce the most detail, and have a clear superiority to the smaller images - which also get smoothed as they are enlarged!

Yep - the math works in reverse too. As a line must occupy more and more pixels, the detail is lost, and lines become more regular in appearance - less natural on irregular shapes, and sharper on more regular ones.

So, lets consider what he did for the "test".

He reduced the 10 and 7 megapixel images to 5 megapixel, thus smoothing them both. If the printer printed at 300dpi, then the resulting raster from 16"x20" was 4800x6000. Thus, after the two higher resolution images were smoothed during reduction, all three were then smoothed during enlargement.

No surprise at all that they looked similar. Mathematically impossible for them to be otherwise.

If he had simply directly enlarged each image directly to the printer, then I would expect the detail to be better on the larger rasters. But looking too careully at linear surfaces might lead you to believe the lowest resolution image was best, since enlargement would tend to sharpen lines.

On the other hand, I suppose that in another sense, all this tends to validate the assertion of the writer - since we must all live with these algorithmic restraints when printing, the inability to preserve detail does tend to make one size raster superior to other sizes - at different print sizes.

I think most people fall down in these concepts, because they still think in terms of optical reduction and enlargement - which tends to preserve detail, because at each size, the original detail is still there, it just occupies more or less space. And so while the accuity may change, the actual detail included does not.

[Bartman01]

I agree with his root assumption, just think that his test is very flawed.

Everything else being equal:
There is not a 'real' difference between a 6 and 8 mp sensor.
There is not a 'real' difference between a 8 and 10 mp sensor.
There is not a 'real' difference betwwen a 10 and 12 mp sensor.

Unless my math is off, it takes a 4x increase in mp count to double the number of pixels in the image.

The problem is that NOTHING is ever kept constant. As an example, the Nikon D100, D70, D50, and now D40 all use the exact same sensor with different results based on the hardware and software surrounding that sensor.

Canon continues to improve their DIGIC (sp?) processing with each new generation of their cameras.

What most people are really 'seeing' are improvements in sensor technology and/or processing. And that is the part of the equation that is being missed in the MP comparison arguments.

[venchka]

In that case, Olympus must have been doing something very right 3-4 years ago because that's when my friend bought the camera that took the 30x40 picture I saw last night.

[jeffkohn]

There is not a 'real' difference between a 6 and 8 mp sensor.
There is not a 'real' difference between a 8 and 10 mp sensor.
There is not a 'real' difference betwwen a 10 and 12 mp sensor.

You could argue that I suppose, although I would substitute 'substantial' for 'real'. But I don't think you can make the leap from those three statements to say that there isn't a meaningful difference between a 6mp and 12mp sensor, or even a 6mp and 10mp sensor.

Unless my math is off, it takes a 4x increase in mp count to double the number of pixels in the image

Your math is off. MP count is a count of the number of pixels in an image, so of course a 2x inrease in MP will give you twice as many pixels. But I know what you were trying to say, it takes a 4x increase in pixels to double your print size. While true, this doesn't mean than anything less than a 4x increase is insignificant or pointless IMHO.

[jdlaing]

Originally Posted by venchka

In that case, Olympus must have been doing something very right 3-4 years ago because that's when my friend bought the camera that took the 30x40 picture I saw last night.

I shoot Olympus DSLRs. And now a Leica.
I'd be happy to show you what it is capable of.
The E-1 is 5 megapixels and the E-330 and Leica are 7.5 each.
It's in the color and other details........not so much megapixels.

The Olympus lenses.........though their prices are not for the faint of heart.......make all the difference in the world.

This "megapixel madness" argument is a long standing one over on some of the Olympus Forums. Some really heated discussions go on over there..........

[PIC]

I use frame size more than mp counts and always think bigger is better.
At 30 x 40 you will see delta between 5 and 13 mp with APS size sensor.
The rule is 2x at the largest print size.
16 x 24 is just not large enough.
You will also see difference between APS and Full Frame DSLR sensor size.

[tragic hero]

Originally Posted by Bartman01

Canon continues to improve their DIGIC (sp?) processing with each new generation of their cameras.

canon is still making improvements to it, the 4mp 1d is not as good as the 8mp 1d, even tho they use the same sensor

[Jake]

Originally Posted by tragic hero

canon is still making improvements to it, the 4mp 1d is not as good as the 8mp 1d, even tho they use the same sensor

They don't use the same sensor.