a Perceptual Mega Pixel explained?

Well and some other thoughts too...

Espeically beginners who have read just a few internet discussion groups before 'buying in' tend to focus on bodies, features and megapixel counts. They ignore lenses and don't grasp the importance of sensor format size. Frequently people buy the best body they can afford and get a crummy zoom. They justify this variously with:

• "I'll get a good lens later"
• "I really want to take good pictures so I bought the best camera"
• "the images look really good" (but I've never actually compared them side by side)

Yet both Lenses and Cameras are needed to take a pictures of higher quality.

Long ago I wrote a WWW post about what I called Megapixel Madness and I thought that it was about time to take that notion further with the new metric which DxO has introduced: the Perceptual Megapixel. In this blog post I thought I'd explore that and provide what I hope is a tangible representation on what it is that I think DxO are on about.

summary of findings

For those who don't have time to read:

• Perceptual Megapixels as a metric seems to make sence in terms of end product image evaluation
• If you are obsessed with every last skerrik of image quality at capture, then you'll need top lenses and >24Mpix Full Frame digital a tripod : or you've got your hand on it.
• if not using Full Frame don't expect much more than effectively (perceptually) 13Megapixels no matter what's written on the camera specs, but in most cases 8
• there isn't much difference between 4/3 and APS-C (gosh, as the maths would indicate)
• its going to cost you a lot to get a lens to get the high Megapixel counts (and some focusing accuracy and probably a tripod too)
• if clarity and contrast are your gig, why aren't you using Full Frame (it may even be cheaper)
• for light weight, compact versatility and access to interchangable lenses NEX and micro43 are hard to beat

going forward

Now first I'd like to introduce a new term, the bajillapixel camera. This if of course dervied from the numeric bajillion (following on somewhere after million....). From Wikitionary

Noun
bajillion (plural bajillions)

1. (slang, hyperbolic) An unspecified large number (of).
   • 2003 27 August, connie, “Tomatoes . . .”, rec.food.preserving, Usenet:

     However, I have a bajillion cherry tomatoes. Is it worth it to try to make spaghetti sauce from them, or will it take 10 bajillion cherry tomatoes to make any quantity of sauce?

So without getting all detailed and accurate on ya I reckon any camera that has more than enough pixels is a bajillapixel camera.

its just a jump to the left..

Beginners (and sadly more experienced photographers too) go on about "image clarity" as it it was something obviously quantifiable. Yet at the same time these folk shy away from metrics - after all its the picture quality which counts right? All this quantative stuff is just hard to grasp and gets in the way of the photographic journey ... then a step to the right. Put your hands on your hips ...

Ok, so too much is never enough ... sure ... well if you're buying the media cards and HDD storage and downsizing for www .... Well anyway, before we get bogged down, what is "more than enough"?

Let me go out on a limb here and say - if the lens system you're using provides the limits to the maximum number of megapixels obtainable then a body that captures more than 1.3 times that is "more than enough".

DxO provides this nifty metric (ok, go get the garlic and crucifixes I said metric) called the Perceptual Megapixel which uses the lens on the camera and gives you a rating. Ok ... sounds great but what does it "look like". To answer that question I have done what I think demonstrates what they are trying to encapsulate in a single number. So presenting my

Virtual Perceptual Megapixel

Fig 1 - workflow

Basically the idea is simpler that it sounds in description, so I hope that the diagram I have to the left here makes it clearer.

Assuming you start with  12 megapixels (or 4000 x 3000) and your camera + lens rating is 7P-Mpix then it would be the same as if you scaled down the image to 7 megapixels (which is 3150 x 2363, and yes the diagram is to scale) and then upscale that image in software (which please recall results in detail loss, if there was fully 12Mpix present, because upscales can not create what is not there, if they could then you could just buy a 1Megapixel camera and upscale all the time).

So lets have a look at this experimentally.

Method

To do this I thought that I would capture a scene which has a (nearly) "Perfect Mexapixel" content. I then scale it back as above and then scale it up as above and compare to an actual capture.

So, how to get a perfect capture? Well stitching a 4 x 4 array of images taken with a x2 focal length (resulting in something like 8000 x 6000) wold do that. For instance, this scene (taken with the Panasonic 14mm f2.5 lens)

when taken with a x2 longer focal length lens, and combined in stitched format (before its properly joined) would have 4 12mpix images:

and would then scale back divided by 2 (to 4000 x 3000) to produce a nearly perfect 12 Megapixel capture.

Since I am going to be pixel peeping major league here I didn't bother with doing all that (well and you can't see a 8000x6000 image on the web anyway) and just took one image with the 14mm and one with the Sigma 30mm f2.8 and work with just them at pixel peeping levels. For the curious I also used the following methods:

• capture in RAW
• used the lenses at f2.8
• illuminate by bounce flash (camera set on manual exposure) to ensure that effective shutter speeds (that would be the flash duration, which would be about 1/5000th of a second) removed any shutter bounce
• mount camera on tripod

So, here we have 100% pixel crops of the outcomes

since that might not look clear enough on these monitors (btw, use the middle mouse button to "center-click" the image to open in a fresh tap at a proper 100% viewing) I thought I'd look at them at 200% in photoshop.

which shows to me that the "virtual perceptual megapixel" still holds a wee bit more detail than the 7PMpix I chose.

This suggests that either I'm wrong in my theory or that the Panasonic 14mm f2.5 isn't as sharp as people whack on about (and that's my experience with this lens btw). As there is no DxO rating that I can find for the Panasonic 14mm I'm left in the 'lurch' on this one for actual data (AKA your guess is as good as mine). However its also possible that the 'perfect' megapixel capture is more perfect than possible because the stitching actually reduced the effect of the anti-alias filter.

None the less it does show the sorts of degradation which are at least what you'll see.

If the theory works then one should be able to scale back the actual capture of the 12Mpix and then scale back up without loss, as there is in theory no extra data there to lose.

So I took the 14mm image and downscaled it to 7MP saved it (as a TIFF) then opened and enlarged it back to 12MP using 'nearest neighbor'. Assuming there was only 7MP (or less) then this "compression" should not create much in the way of loss

So, in blind testing, you tell me: which is which?

I have one further point to add on the creation of the "perfect" 12Megapixels and that is the the role of the anti-alias filter will be reduced (making it a little less soft) than if you had a way of perfectly getting light to the sensor in a 12Megapixel manner.

Discussion

So were you surprised?
Did all the changes  (looking at the above Fig 1 of workflow) result in as much change in clarity as you may have expected? I suspect not.

To me this shows that the differences between 7 and 12Mpix are perhaps less than people appreciate (something I've been banging on about for a while) and that to see really significant differences you need to go up more than x2 in dimensions, meaning that you need a good 36 megapixels to see any substantial increase in clarity. I'd call that diminishing returns.

Now think about that for a while. How big is that file going to be, how large will it be in memory, how long to demosiac how much to store? The answer is basically x4 ... because to see a x2 improvement in image detail you need x2 on each dimension (as well as the lens systems able to capture it).

To be honest, how often are you going to actually print anything big? I know from personal experience that I have taken images that were 4000x3000 and had them printed at 150dpi will yeild a print that is 68 x 50 cm or 27 x 20 inches. Do you really want your camera to be capturing 8000x6000? Thats 48Megapixels.


With what we've seen above lets look at a few Lens / Camera combinations in the DxO listings.

Some top rated micro4/3 lenses

Nothing higher than 9P-Mpix, and that is a prime. In the zooms (which people froth on about as being 'stellar') are quite limp in reality at 7 and 8.

For those smug "well I've got APS-C not that winky dink 4/3 sensor" folks let me present that as a compare:

So on a camera like the 18Megapixle 7D your total capture still only gives you 9PMpix ... looks bad even with one eye to me. This backs my view that back about the 12Mpix times on the APS cameras we were hitting limits

The problem is actually capture size.  Bigger is Better in that area (and I mean area not just cutting your pizza into more slices), a well known point among photographers with any training. Lets look at some full frame examples. I chose the EF 35mm f2 lens as I also used that to look at the 7D above:

So the trusty old 5D 12Mpix full frame gives 11 P-Mpix (nearly a perfect match with the sensor) and there just isn't a lens combo that gives you this IQ in APS even though the bodies are mostly the same size.

Even with Full Frame, its a diminishing return as each higher MPix cound gives progressively gives less Perceptual image quality. So as the camera M-Pix goes up to 21Mpix the end result is still only going up to 17P-Mpix. Diminshing returns to me, unelss you are willing to pop some BIG BUCK$ on some Zeiss lenses.

I'll leave it up to the reader to determine the ratio of dollars / P-Mpix for you to work out if you (like me) would be happier with a $500 full frame that gets 80% of the image quality of a $2000 body and some $5000 lenses.

To drive this point in (*of format size being the most important factor for end result IQ), I thought I'd compare a EOS 40D (10 MPix APS camera) with a couple of Full Frame cameras with various lenses.

So with a lens that's much better than the reglarly purchased "kit" lens (meaning on a lesser lens you'll get less P-Mpix) the 40D is down from its real10MPix to 6P-Mpix and the 36MegaPixel Nikon D800 only gives its best with a uber expensive Carl Zeiss lens (and giving a piss poor 13P-Mpix with an expensive Nikon lens that people will probably go 'ooooh' over).

My point here is that if you are going to get a Full Frame camera with a high megapixel count then you'll need to back it up with actually stellar lenses. So the DxO P-Mpix rating actually gives you a good tool  to help in that decision making process. My other point is there is little point in getting a APS or 4/3 camera sensor in much higher than 16Mpix because you're not going to get any benefit.

This suggests that in reality (something I've argued for a while) that a good 6Megapixel image is really quite enough for most things and certainly more than we need for our big screen HDTV's our phones or our tablets. These 12Megapixel captures (on either of the smaller sensor formats) are showing Perceptual Megapixel values of 9P-Megapixels and that's not really far from a really good 6Megapixel.

Chasing IQ

So far I have not even touched on the other significant issue, and that is of camera shake. Even IF you have a potential for 28P-Mpix on the Nikon system (or on some other system) you just won't get that high IQ without the camera on a tripod and attention to things like shutter bounce. So its quite possible that your hand held images are in reality back down to 12P-Mpix if done hand held or with a crummy tripod. Even tripods are not all equal as I discovered some time ago with problems I had with my (then) new Manfrotto 190.

Basically as soon as you start humping a good tripod about you can kiss goodbye to any weight or size savings in the camera body. So if you're after image clarity as a goal you really need full frame and a tripod.

This raises the ugly question of "if you're going to hand hold, why the hell would you hump around bloated DSLR cameras which only have APS sensors in them?" The evidence is that you won't get observable IQ differences (refer again to the above images) using such cameras.

This is exactly the reason I dumped my EOS 20D, didn't go further in EOS APS and got into micro43.

To me the only reason to hump around a bloated APS sensor camera like a Canon 70D is because you're into the feel of a heavy camera. There is perhaps also the argument to be made of better AF speed on telephoto work (read sports) and being able to get better 'telephoto effect' from the same telephoto lens as compared to Full Frame.

To me unless you're using a tripod and care, then cameras like the NEX and the micro43 cameras yeild all the image quality with significantly less weight and so unless there is some utility to be gained from the brick in your hands, you will be better off with Full Frame (especially perhaps with a Sony A7) than a bloated APS camera.

So hopefully this new metric from DxO will allow photographers to plan what they need and make better informed purchases of both camera bodies and lenses. Wangers of course won't be worried about how the images look because they're more worried about being seen with the right gear ;-)

A classic from WhatTheDuck to sum up

what I'd like to see as outcomes

Based on these findings I would love to see (as I have since 2001 when disappointed with the EOS D30) more of a push for larger sensors (full frame) but with no major emphasis on pixel density (12Mp is enough for most) and instead a greater emphasis on actually giving better bit depth. The current RAW formats are under used and even adding another most significant bit depth to the data would make for better highlight handling and enable us therefore to get better shadows (by exposing to the right).

This analysis makes it clear to me that not only do larger sensors do a better job of getting Higher Image Quality (IQ) in terms of effective pixels but all my previous experience shows that larger sensors also give better contrast and bokeh.
:-)