19
\$\begingroup\$

The question is inspired by this question showing these pictures.
The accepted answer suggests that these pictures were taken by a 8x10 view camera, and the use of a 8x10 camera was confirmed in the comments.

My question is: How can you tell?


When viewed on the webpage these images are 496x620 = 0.37 megapixels (or 720x900 = 0.65 megapixels if you click for "full view").
So any camera with a resolution higher than 0.37 Mpx should be able to capture these pictures, meaning pretty much every smartphone and webcam on the market.

I know about Bayer sensors. But the worst-case effect of a Bayer sensor should be to reduce resolution by a factor of four: If you downscale the picture by a factor of two in each direction, each output pixel will contain data from at least one input sensel for each of the R/G/B channels.
Downscaling by factor 4 still means than any camera with more than 1.5Mpx resolution (rather than the 0.37Mpx of the output) should be able to capture these pictures. We're still talking about pretty much every smartphone and most webcams on the market.

I know about color depth. But JPG, the format we are using to view these pictures, is 8x3=24 bits. And according to the DxOMark scores there are several cameras, including the Sony NEX 7 and Nikon D3200, that are capable of capturing 24 bits of color.
So even if a $10 webcam can't quite capture the nuances in these pictures, a NEX 7 or D3200 should be able to do so.

I know that most lenses have lower resolution than what most sensors are capable of. As an example, the Nikkor 85mm f/1.4G is Nikon's sharpest lens according to DxOMark, and gives a best-case equivalent of 19Mpx resolution on a 24Mpx camera (the full-frame Nikon D3X), while the least-sharp lens has a best-case equivalent of 8Mpx on the same camera.
But the worst lens in their database still gives an order of magnitude more resolution than the output format of these examples.

I know about dynamic range. But these images control the lighting so they neither blow the highlights nor lose the shadows. As long as you are in a position to do that, dynamic range doesn't matter; it will be mapped to the 0-255 output range of JPG anyhow.
In either case, DxOMark says that several cameras with full frame or smaller sensors have a better dynamic range than the best of the medium format cameras.


That's what I know, and there is nothing in these fragments of theory that can tell me how it is possible to tell a 8x10 view camera from a Sony NEX 7 when you view the result as a 0.37 Mpx JPG.

Essentially, as far as I understand, it should be irrelevant how many megapixels and how much color depth the sensor can capture, as long as it's at least as much as the output format can represent.

Still, I don't doubt the judgement of the answer from Stan Rogers. And I've never seen anything similar, in terms of perceived sharpness, from small-sensor cameras.

Have I misunderstood what resolution means?

I guess I'm primarily asking about theory: How can a difference between two resolutions (measured in pixels, lp/mm, color depth or whatever) be visible in a display format that has less resolution that either of the originals?

Or to phrase it differently: Is there anything to stop me, in principle, from replicating these pictures down to the pixel by using a Sony NEX 7 and $10,000 worth of lighting?

\$\endgroup\$
2
  • 2
    \$\begingroup\$ As long as you get everything perfect in your shot there isn't much advantage to shooting RAW over JPEG either. Resolution, bit depth and RAW data is all about getting the most from your photo! Small web photo today, 4x retina display or high-quality print tomorrow. \$\endgroup\$
    – Henrik
    Jan 9, 2013 at 7:50
  • 1
    \$\begingroup\$ Unfortunately I can only accept one answer. Matt Grum makes a credible argument that a small sensor, plus lighting and a bit of post processing, can replicate perhaps 95% of what we see in the Martin Schoeller pictures at web size (excluding the DoF effects). I think Michael Nielsen makes a credible argument that you might need better than typical APS-C system resolution to get to 100%. And Stan Rogers explains how we can recognize a large format camera even in web sized photos. I suggest you read them all :) \$\endgroup\$ Jan 11, 2013 at 10:56

4 Answers 4

12
\$\begingroup\$

It's all about the micro contrast. Look at the posts about aps-c versus full frame and then extend that difference to medium and large format sensors.

When do the differences between APS-C and full frame sensors matter, and why?

Following the theories on oversampling, it is better to sample at a higher rate and then downsample than to sample at the nyquist limit from the start - ie. if your end goal is 640x480 , it is still better to use a 1280x960 sensor than a 640x480 sensor.

It doesn't matter how many MPixels you have when neighboring pixels depend on each other, anyway, due to the circle of confusion being larger than your pixels on the sensor plane. The lenses have limited ability to resolve, too. Furthermore, you have to consider the lens "sharpness" versus its aperture, and a larger sensor allows you to get closer and get narrower DOF stopped down, which means you can capture even more details - Circle of confusion is larger, lens is working with less diffusion, etc.

And then you have the "depth compression" done by the focal length of the lens that is pretty aggressive in those shots, pointing to a telephoto. The FOV on a small sensor would require you to step back a long way and open up the aperture a lot to get that narrow DOF. However, running the numbers, with a full frame camera you could achieve it, 210mm, 2meters distance, F8 would give a 4 cm DOF and a FOV that takes just the face like those shots.

Put in another way: the larger the sensor relative to the subject, the less the lens has to work on the light rays to compress them into a tight spot. This increases the clarity of the shot and it shows no matter the viewing distance (which is what is being simulated by resizing the image to lower resolution).

Folllowing discussions about detail enhancement and retention through resizing here's a comparison if similar subjects large format vs fullframe and large format versus apsc:

Top: male faces with beard stubs. In the resolution on the site you link to, the beard is rendered with pixelwide hairs, but all that is lost at the same size as Matt's example. Now the beards are diffuse. If we see Matt's image in the same size as the 8x10 photos in the site, we might see a big difference if the head isnt in focus. Even a aps-c system and smaller sensor could produce this result (regarding details).

Bottom: is we compare the female face eyelashes at similar size as it is on the webpage you showed, to a in-focus eye from a aps-c camera, and sharpening is not going to bring the pores in the skin back. We might enhance the perception of the eyelashes at the cost of a bright halo around it.

We now see a huge "overall system" resolution difference, and the apsc camera + the lens used + seen at the given lowres resolution cannot render the same detail as the 8x10 camera + that lens + the viewed resolution could. Hope my point is clearer now.

details

Another comparison to aps-c, beard stubs, after sharpening them. Even though stackexchange resizes them we can still perceive a difference in clarity.

apsc

In conclusion the other factors you are asking about other than the pixel resolution are:

  • Total system resolution lp/mm
  • SNR
  • Magnification from the person to the sensor to the screen you view it on at the given distance you view it from to your eye's projection on the retina resolution - the smaller (below 1:1) magnification any part of the system , the higher the demands for the above two factors, which in turn are negatively influenced by the smaller projection area.

You'll get more details in a downscaled macro shot than you do without shooting macro in the first place.

A final proof that resolution before downscaling matters.Top: 21MP FF Bottom: 15MP Aps-c with the same lens/aperture focal length.

Two diff sensor res/size

Now rescaled to equal resolution:

Small

Smaller

Smallest

Tiny

and applied a bit of sharpening to bring back some detail. What do you see? a bit more details from the 21mp FF camera viewed at the same size/resolution which would be equivalent down to a 3Mp camera. you can't count the lines in the rescaled image, but the perception that they are lines is true. Whether you want this or not is your creative choice, but starting with the higher resolution (given by the total system) you get the choice. If you dont want them you can blur the image before rescaling.

One final experiment to show the difference between a small size, low res vs larger sensor, higher resolution , but rescaled and sharpened to the same resolution, shown at the SAME SIZE in the end - with ALL ELSE EQUAL. Cool, eh? How did I do that? My APS-C camera I take simulate a "crop sensor" (smaller than my apc-c) by cropping an image out of the image. Then I go closer to the subject to fill a 4x larger sensor with the same subject. - like the portraits on large format sensor is basically a macro shot - getting much closer than you'd get with a aps-c camera. Same electronics quality, same lens, same settings, same light.

This is how it looks on the small sensor , let's call it "mini aps-cc":

Small sensor

Here we see the "large format" (large full aps-c ):

aps-c

Here we see loads of details, right? But that doesnt matter after we rescale it down to a 0.016MP image, and sharpen for overall same contrast, does it?

Comparison equal size

But indeed we do! If you still don't believe me, I give up :)

\$\endgroup\$
7
  • \$\begingroup\$ A note about the maze pattern these crops display: it's the bayer interpolation artifact common for AHD bayer demosaicking. \$\endgroup\$ Jan 9, 2013 at 16:46
  • \$\begingroup\$ However, the perception that it is an array of lines is more true to the original than seeing a grey box. Same as seeing pores in skin , maybe not the exact amount as real life, but they'd be there. \$\endgroup\$ Jan 9, 2013 at 17:06
  • \$\begingroup\$ Thanks, interesting comparisons. It does look like sharpness/detail (or at least an illusion of detail) survives more downscaling than I thought. You've given me something to consider. \$\endgroup\$ Jan 9, 2013 at 18:03
  • \$\begingroup\$ keep in mind that converging lines is a nightmare for interpolation algorithms and we KNOW how it SHOULD look, we dont have the same assumptions about the details in a nice skin texture, so even is a spot moves a half pixel in the face it still looks awesome. In the smallest example above, we still perceive that it has lines down to the second vertical. \$\endgroup\$ Jan 10, 2013 at 19:07
  • \$\begingroup\$ Michael, a related question -- what if the sensor size is the same? Am I going to get better results from an APS-C sensor with a 16 MP native resolution, but downsampled to 8MP, say, as opposed to buying a (hypothetical) camera with an 8MP ASP-C sensor? \$\endgroup\$ Dec 13, 2013 at 4:04
8
\$\begingroup\$

As much as anything, the view camera give-away was an extremely shallow depth of field that is not parallel to the film plane. Getting the lips and eyes in focus in several of the images with the posture of the subject and the very shallow DoF (as often as not, the forehead in jawline immediately below the eyes is soft) is impossible without tilt.

There's also the nature of the out-of-focus ranges; it's more typical of near-life-size rendition than far focus. The DoF extension toward the camera is nearly equal to the extension toward the background, which is what you'd expect as you approach 1:1 reproduction. At more normal working distances, you'd expect about a one-third/two-thirds distribution of apparent sharpness around the plane of sharp focus (one-third of the DoF in front of the plane of focus; two-thirds behind). So that indicates a "sensor" considerably larger than APS-C, full-frame or even medium format. Experience with the 4x5 and 8x10 formats told me that it was more likely to be an 8x10 (although even one of the rare larger-format cameras would not have been out of the question either), and I thought the 210mm lens more probable than, say, a 150mm based on perspective (which is tight, but not quite claustrophobic).

As Michael Nielsen pointed out, there is a lot of "micro contrast" apparent in the images, but to a degree you can fake that in post-processing, especially if the aim is to render at webby sizes. And I suppose you could even fake the DoF and focal plane if you were diligent in creating a depth map with a gradient focal plane applied against what would pretty much have to be a 3D model of the subject and understood the focus dynamics of an 8x10 at about 50-60% life size reproduction, but that would be one heck of a lot of work. The economical solution, both for the photographer and anyone analyzing the image, would be an actual 8x10 view camera.

\$\endgroup\$
12
  • \$\begingroup\$ Increasing the microcontrast in post introduces artifacts, though, as well as the algorithms are "guessing" the details, so it wouldn't be the same picture. \$\endgroup\$ Jan 9, 2013 at 8:35
  • \$\begingroup\$ @Michael but in a resized reproduction those could be for all purposes not apparent? Or not? \$\endgroup\$
    – Francesco
    Jan 9, 2013 at 8:49
  • \$\begingroup\$ Was "to a degree" ambiguous? \$\endgroup\$
    – user2719
    Jan 9, 2013 at 9:27
  • \$\begingroup\$ Do you mean if you sharpen first and then resize where the enhancement gets wiped out, unless the highpass filter window is so huge that it introduces ringing across the frame? The details in those portraits are so incredible, that if you examine a 1x50 pixels crop of the skin, it looks like a binary image, which can easily produce halo rings with sharpening filters. \$\endgroup\$ Jan 9, 2013 at 10:23
  • 3
    \$\begingroup\$ @MichaelNielsen when doing a huge downsample you can introduce aliasing in a totally soft image (with no microcontrast whatsoever) and get an extremely crisp looking image with no halos or other sharpening artifacts. If it wasn't for the tilted plane of focus you wouldn't be able to tell those images were large format by looking at the apparent "details" \$\endgroup\$
    – Matt Grum
    Jan 9, 2013 at 11:36
5
\$\begingroup\$

No, a high resolution camera really doesn't matter when the output is low resolution, at least not when you go from double digit megapixels down to a quarter of one megapixel. Take the following image:

Resized for the web it looks fine despite the fact that the subject's face wasn't even in focus! It's obvious when viewed at 100% and obvious in print, but you absolutely cannot tell by the time it's been resized to 2% of it's original pixels and sharpened for the web.

Here's another example, take a massively soft original image, at about 8 megapixels:

Heavily downsample to a web friendly resolution and sharpen, and suddenly look at all the microcontrast!

A 2x oversampling will definitely help with the resolution and colour fidelity of Bayer images. But an image from the original Canon Digital Rebel (300D) released way back in 2003 resized to 600x400 is a 5x oversampling, in each direction meaning each pixel in the resized image takes the place of 25 original pixels. Very, very little of the quality of those 25 pixels is going to make an impact on the resized image.

The increased microcontrast afforded by a larger format system simply wont be visible, the macrocontrast that you do see can be made up for by postprocessing, when you've got so much resolution to throw away the sharpening artifacts wont be visible.

If you matched the depth of field then it would be extremely difficult to tell the difference between a 10x8 view camera and a compact when resized to less than 1 megapixel.

\$\endgroup\$
9
  • \$\begingroup\$ can I see his face in the same size as the portraits? \$\endgroup\$ Jan 9, 2013 at 12:48
  • \$\begingroup\$ yes when I get home and dig out the originals! \$\endgroup\$
    – Matt Grum
    Jan 9, 2013 at 13:04
  • \$\begingroup\$ I have to disagree with the blunt statement "a high resolution camera really doesn't matter when the output is low resolution". Oversampling and magnification ratios do matter in the final results (to a certain limit). \$\endgroup\$ Jan 9, 2013 at 13:22
  • 1
    \$\begingroup\$ you are getting nice details in the small image exactly because you oversampled so many times. If you had used only 2x oversampling with that same amount of blur wouldn't have gotten any of the skin details. And you overdid it with sharpening, his hair looks weird. Furthermore, this shot looks like a typical low end image. nothing like those images we are talking about, that retains more details than a APS-C system with a decent lens can produce. \$\endgroup\$ Jan 10, 2013 at 17:35
  • \$\begingroup\$ Thanks for the latest example, it makes it seem credible that something like the pictures linked in the question could be created with an APS-C sensor. @MichaelNielsen But that's the kind of oversampling the question is about, e.g. downscaling say a 16Mpx APS-C shot to <1Mpx web size. So in terms of oversampling, can e.g. 7x (49 pixels input to 1 px output) oversampling be visibly better than 5x oversampling? I'm leaning towards "probably not", in your example with the lines I can't see much/any difference beyond 4x oversampling. \$\endgroup\$ Jan 10, 2013 at 18:28
1
\$\begingroup\$

When rendering a scene with fine details into a low-resolution image, it is necessary to apply spacial filtering to remove any content whose frequency is above the Nyquist limit. When applying spacial filtering, there are two conflicting goals:

  1. Content whose spacial frequency is low enough to be shown should be attenuated as little as possible.

  2. Content whose spacial frequencies are similar should be attenuated by roughly-equal amounts.

To see where the goals conflict, imagine a scene with a pattern of converging lines. If the spacing gets somewhat near the Nyquist limit for the target resolution, but the lines are always separated enough to be shown clearly, it will often be better to show them clearly than to blur them. If, however, the scene contains a pattern of converging lines which get too close to be distinguishable, having the lines become gradually more blurry as their spacing approaches the resolution limit would be less distracting than having the lines appear clearly up to a point where they sharply transition to a solid gray.

In many cases, the optimal kind of filtering for a scene will depend upon its content and which aspects of it are of interest. Capturing a scene in a resolution higher than its intended output format will ensure that all information which might be desired in the final image will be retained. Information will have to be filtered out before the image can be rendered in a lower resolution, but the high-resolution capture will make it possible to tailor the exact filtering methods to optimally fit the needs of the scene (and possibly use different methods for different parts of a scene). If one is intending to render a final image at 640x480 the benefit of capturing it at 6400x4800 would probably not be much greater than that afforded by 1600x1200, but there may be some benefits to going up to about 2.5x.

\$\endgroup\$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

Not the answer you're looking for? Browse other questions tagged or ask your own question.