Red One Measured Resolution

[Graeme Nattress]

Yes, that is the goal, and yes, I think we succeed at it. One thing people forget about aliasing is that it confuses MPEG2 and other motion estimation codecs, and excessive edge sharpness can lead to greater perception of low fps motion stutter. One of the reasons that film can "get away" with 24fps is that it's usually shot with shallow dof, and it doesn't have especially enhanced edges.

Graeme

 

[Michael Lindsay]

There is NO doubt in my mind that RED image is unique in that it pushes more than enough resolution without the usual "electronic" look.

I'm so thankful..

The worry is that people are confused by the cheap fix that is aggressive edge enhancement. Edge enhancement is MSG for pictures.

I am so excited about the 'possibility' of working without any edge enhancement. Unless as a a fix or to achieve a specific look.

Michael Lindsay

 

[Dj Joofa]

So, I do think it's good to start with a high number of samples, to allow you to use proper filtering, and go for a smooth, continuous, nice image reconstruction, and not worry too much about absolute resolution, because you started with enough. It's when you start with what you want to try to end up with, that you get "into trouble".

Graeme

True, fully agree with that.

(A passing comment on what we are talking about before. In case you have not had a chance to go over Mertz and Gray classic 1934 paper, please do it -- I think that paper is a must read for anybody trying to understand TV scanning process.

Ref.: Pierre Mertz and Frank Gray, "A Theory of Scanning and Its Relation to the Characteristics of the Transmitted Signal in Telephotography and Television," Bell System Technical Journal, Vol. 13, pp. 464-515, July, 1934)

 

[Dan Hudgins]

4k and lenses.

4k and lenses.

Oh okay, thanks for the information.

Did you try gray scale charts, or red and blue colored charts were also tried? If it is not considered internal or private information, can you kindly inform what lens was used.

If you use a 35mm movie camera lens that covers full 18.36mm height aperture, very few lenses can resolve better than 30lp/mm full open, or 50lp/mm at best stop at 80%MTF. In the corners the numbers might be 25 lp/mm for a f2.0 lens to 10 lp/mm for a f1.2 lens.

That works out to maybe 30lp/mm x 18mm = 1080 raster lines, about.

To get 2048 raster lines would require a lens that can make 57lp/mm at 80% MTF.

To get 4096 pixels high image would require a lens that can make 114lp/mm at 80% MTF, something that would be almost impossible in a fast camera lens.

To get better results from the Bayer filter you want the circle of confusion of the camera lens to be two pixels wide, so if you shoot a red/blue checkerboard chart the squares would need to be two pixels wide or bigger, three might resolve, so each patch would be 9 pixels on the sensor.

Using a sensor 8k wide is probably the point of diminishing returns with regard to resolution in 4k output since high speed camera lenses cannot have enough of the frame in focus on the sensor to record details that small.

If you make extra sharp camera lenses with max stop of about f/4.5 to use on the 4k sensor you would just make the Aliasing and Bayer Artifacts worse, unless you go to a 8k sensor, where each four pixel group maps to one 4k pixel so no de-Bayer is needed (although might help a little.)

To check the "resolution" of your camera there is a simple test, at night put a flashlight lamp with a small filament about 100 yards away, this should be a clear bulb without anything around it. Focus the camera on it with a good semi-wide lens like a 35mm. Shoot some frames. To get the compression working you might want some of the surroundings lit to use bandwidth. Then look at the filament spot in the image, which should be one pixel (you might need to pan the camera slowly to get the light to sit on a single pixels center), using graphics program that lets you zoom in 1000% or so so you can see the square pixels. You can also try using colored filters over the camera to see what happens to the point image in colors. Do not over expose too much since the exposure can make the image larger or smaller.

A variation on that is to poke two pin holes in some black wrap and light from behind with a diffuser. The holes can be about 1/10th the distance between them. You can then take images at various distances until you can no longer tell that there are two holes (with an equal amount of back between?), and figure the resolution from that.

At 4k on 35mm color negative film single point objects will scan out to maybe at about 9 to 25 pixel blocks at 10% signal as the radius and may be larger with fast lenses. Getting only a single pixel to have accurate RGB of the right value at 4k is not something you get with a normal camera lens. Using the same lenses on the sensor with OLPF might be a little better but still probably not smaller than 9 if you go down to 10% signal as the radius.

I do not know if the compression would block the pixels around the bright one to make the readings look better than they are, maybe someone could shoot the flashlight bulb and two holes test and post some results? The images need to be gamma-ed up to see the full radius of the spot image.

 

[Dj Joofa]

If you use a 35mm movie camera lens that covers full 18.36mm height aperture, very few lenses can resolve better than 30lp/mm full open, or 50lp/mm at best stop at 80%MTF. In the corners the numbers might be 25 lp/mm for a f2.0 lens to 10 lp/mm for a f1.2 lens.

DanCad3D thanks for the information. I agree that lens MTF effects can cause all sort of problems. Therefore, ideally the sensor response for resolution measurement should be done independent of lenses. And, this is a process that is many times employed in camera manufacturing places -- sensor is exposed to light in specially created hardware kits without the lens. I do not know why Red is using a full-fledge camera (with the lens) to measure resolution.

The reason I was asking for red/blue lines for resolution tests was that a grayscale line will extract responses from all red/green/blue channels of the Bayer grid. So even at the limit of Bayer grid where the line pair distance has become small enough to be in comparison with the separation between red/green/blue photosites, the grayscale lines may still be reconstructed to some extent. However, a red or blue line, say a blue line, will have more trouble if it is falling only on the red-green photosites and not on the blue-green ones, and therefore, may be more difficult to resolve than grayscale lines.

 

[David Mullen ASC]

How does digital relate to film, where larger negatives (65mm, IMAX) need less enlarging to fill a theater screen, and thus can compensate for some limitations of the lens? For example, I've seen 65mm tests shot with older lenses and the image still outresolves 35mm shot with the best modern lens. This is due to the degree of enlargement being smaller.

Would there be a digital equivalent to this -- i.e. could a 8K sensor that is 65mm wide produce a higher resolution image using a lens of the same MTF characteristics as one on a 4K sensor camera that is 35mm wide? Is the image "size" the size of the sensor or the size of the digital file in terms of horizontal and vertical pixels?

What if you made camera that had a 65mm wide sensor and one with a 35mm wide sensor, both 4K (so the photosites on the 65mm version would be twice as large I guess) -- would the physically larger sensor have any image improvements? Any better resolving capability? In the DSLR world, what would be the differences between a 12MP sensor that was the same size as the 35mm still frame versus a 12MP sensor that was half that size, like the 35mm cine frame?

 

[Graeme Nattress]

True, fully agree with that.

(A passing comment on what we are talking about before. In case you have not had a chance to go over Mertz and Gray classic 1934 paper, please do it -- I think that paper is a must read for anybody trying to understand TV scanning process.

Ref.: Pierre Mertz and Frank Gray, "A Theory of Scanning and Its Relation to the Characteristics of the Transmitted Signal in Telephotography and Television," Bell System Technical Journal, Vol. 13, pp. 464-515, July, 1934)

If you have an online link for that paper, that'd be appreciated.

Graeme

 

[Graeme Nattress]

Why a lens for rez tests? Because it gives a real world practical measurement. A camera without a lens is just a funky electronic box with some light sensitivity

Graeme

 

[Dj Joofa]

If you have an online link for that paper, that'd be appreciated.

Graeme

Graeme, despite my best effort I could not find an online link that goes all the way back to 1934. Bell Systems Technical Journal is not part of IEEE, and I don't know when that will be made available online.

I do, however, have the printed copy of the paper, and I can certainly mail you a copy if you pm me an address.

 

[Dan Hudgins]

Is bigger better?

Is bigger better?

How does digital relate to film, where larger negatives (65mm, IMAX) need less enlarging to fill a theater screen, and thus can compensate for some limitations of the lens?

The point I was making was that 35mm movie camera lenses do not and did not need to resolve single pixels 4k (scan lines) size with high MTF like 80%, that is most of the light falling inside a circle one pixel in size, and that on a Bayer sensor that is good since if the lens was that sharp and you took the OLPF out so that you could make a MONOCHROME resolution test reading close to 2000lp the Aliasing would be worse. If you color balance the camera with filters and shoot a gray card, the un-de-Bayer image should show no checkerboard, so except for the fixed pattern noise added by the Bayer filter, the MONOCHROME image should have about the same resolution as a MONOCHROME camera. So if you put a 35mm movie camera lens on the RED ONE (tm) camera with the OLPF in place and shoot a point light source smaller than one pixel by about 10x, you can see the actual resolution of the camera lens, OLPF, micro lens array, cover glass, and such by looking at the spot size in the frame shot, which will not be down to a single pixel like it would be in a high resolution laser film scanner if one could be built to scan IMAX size film and have a spot size small enough.

IMAX size film might be able to hold a 4k size pixel in the mid-tones in the cyan layer, but in the over exposed areas the dyes will bleed, and in the shadows the grain will mix with the details. In the yellow dye the image may not get much past 2k even at IMAX size image, and has more grain.

In lenses of the same design the circle of confusion is proportional to the focal length (a 100mm would be half as sharp as a 50mm of the same perscription scaled) except for some things:

1) The manfacturing tolerances for a 65mm lens are smaller than for a 35mm lens in proportion to the image width, so if the manfacture can center the lenses to 0.001mm the 65mm size lens will be better assembled in relation to the image width, and so will look a little sharper even if the optical prescription is the same.

2) Optical glass has uniformity issues, so getting a bigger piece of glass to make a 65mm lens from that will be better than a 35mm lens will cost much more, when making the smaller lens good parts of the glass can be cut out, so if the lens manfacture does not inspect the glass the 65mm lenses may have more individual variations.

3) When lenses are made for 65mm size images such as 2 1/4 they are usually redesigined to have a slower max f stop, like f/3.5, using the same number of elements the circle of confusion can be made smaller in relation to the image width. If you try to make lower f number lenses for 35mm size images you run against manfacturing tolerances so the lens for 35mm will not be as sharp as the 65mm lens in relation to image width unless the optical shop works to have the 35mm lens made to "super human" tolerances. Making a 35mm lens that gets 150lp/mm is harder than making a 65mm lens that gets 75lp/mm, for the most part.

4) If you shoot with a 65mm lens you get more DOF issues when you are closer than maybe 6 feet than you would with a 35mm lens since the 65mm FL lenses are a larger part of the distance to the subject. Since the 65mm lens will have less DOF on close-ups you would need to stop down more, and if you stop down too much Defraction will make the circle of confusion larger.

When shooting on 35mm film, the lens is almost always sharper than the film being used at stops between f/8 and f/4 (despite the figures the film manfactures dream up using tests that do not test the film at print contrast for the grain and bleed), you can see more detail on the ground glass in the 35mm camera than you can see on the processed and printed film, almost all the time, in 35mm.

At 65mm the lens and the film may be close in resolution, maybe, with the lens maybe a little better at best stop and a little worse wide open.

At IMAX size the film might have a slight edge in the cyan or magenta layer depending on the stock and lens used.

On a digital sensor the dye bleed problems are not there, the OLPF may make the light flair a little more but probably is more even in the highlights and shadows. In the RED ONE (tm) using the same lens as on 35mm the resolution is probably a little better in the 4k image than in a 35mm film print, but that resolution is not near a 4k computer animated image because the lens and OLPF are making the image points larger than one pixel.

Because small lenses cannot resolve small pixels on small sensors a camera with a small sensor may show less Aliasing and Bayer artifact than a large sensor. If you use a 4k IMAX size sensor to get the lens to resolve each pixel well, that would just make the Aliasing and Bayer artifacts worse unless you have a heavy OLPF, so you do not get much ahead, unless you make the IMAX size sensor have the same small pixels, like 16k, so you can use a OLPF with a smaller radius in relation to the image width, then you might get a true 4k image, i.e. like a tri-color drum scan from an 8x10 negative or a computer generated image.

In the older Cooke Series 1 lenses or Baltars they all used 6 elements and had different focal lengths, so the 24mm or 28mm would in general be sharper in the center than the 100mm because the circle of confusion was in part in relation to the focal length. The longer lenses were re-optimised to have a smaller coverage to in part make up for the longer focal length and get back some of the lost sharpness. A 35mm 6 element prime lens at f/6.3 is probably as sharp as you can get in the center of the frame for a 35mm format lens, and that still would make a spot larger than a single 4k pixel.

To cut in 4k CGI images with the 4k RED ONE (tm) images the CGI would need to be fuzzed up or they would look like a mismatch, like in a plate you would not want the plate sharper than the foreground.

It might be possable to make a lens for the RED ONE (tm) with many elements at high cost that could resolve 4k size pixels, but since the camera cannot resolve 4k size pixels it would probably make things look worse. I have had this problem scanning film, with the scanner lens set to the "best" stop you get all kinds of cracks, marks, and krud showing up, not to mention very well defined grain. So I have had to stop the lens in the scanner down to increase the scanner circle of confusing to at least half the camera lens circle of confusion in order to reduce the grain pickup and artifacts between the film grain and the sensor pixels.

Because of the problems with projecting images at 4k resolution I am not sure any projector can get 80% of the light in a single pixel on the screen. You can test your 4k projector by making a black and white checker board in single pixels and see if the contrast is 80% by moving a light probe on the screen. To work out peaking, make test patterns that have black and white single pixel patterns of horz and vert lines also and probe those. You can also make a "stair" field" of single pixel light gray spots or colored spots on a dark gray background, and dark single pixel spots on a light background and measure the spot size on the screen with a mm gauge, then divide by the image width or height to get the actual resolution. It should be very hard to focus a projector well enough, or to make a lens good enough at f/2.0 or faster, to get 4k at 80% MTF on a screen 20 meters wide, especially in the corners. You could make an IMAX test film by shooting IMAX print stock in the IMAX camera with an animation motor, the cyan layer in the print stock should be able to hold spots that single 4k pixel small, and see if Digital 4k projection is better or worse than IMAX. The lens used in IMAX projection is probably the weak link, it could be fixed with a stop, but the image would be too dim then.

Another reason that a 70mm print projected looks better than 35mm projected is that the back focus tolerance is about the same, but the 35mm lens will look more out of focus because it is a shorter focal length and magnifies the focus errors caused by the print not being flat in the projector gate. A 70mm print of a 35mm negative would look better in most cases than a 35mm print projected. Also the 35mm film limits the resolution in the negative and print, whereas the 70mm projection lens limits the resolution in the 70mm projection, for the most part. The projection lens is almost always worse than the camera lens, and being a little out of focus makes the whole frame out of focus, whereas if the camera lens is out of focus only part of the frame is out of focus, that is something is in focus just off the subject a little.

The OLPF is like a projector being a little out of focus, it puts the whole image soft. If you shoot without the OLPF only the parts that are in sharp focus will show very bad Aliasing and Bayer artifacts if the lens circle of confusion is soft on the edges. Some lenses make a bright ring at the outside of the circle of confusion and may show Aliasing and Bayer artifacts even in out of focus parts of the frame, that may be why some Japanese have been putting money into research on lens "Bokeh", the OLPF needs to dampen the "Bokeh" edges to reduce the bandwith eating overhead of the JPEG2000 or other compression.

http://en.wikipedia.org/wiki/Bokeh

Lenses for DSLR cameras need to have a "soft" focus, whereas movie camera lenses wanted the best fine point the maker could get so some fine detail could get into the surface of the film and show up as low % MTF on the screen since eyes can see low % MTF in film images, but in digital images those low MTF just make artifacts with the fixed size pixels and make the compression work harder.

 

[Hans von Sonntag]

Good reading!

Thanks,

Hans

 

[Blair S. Paulsen]

Food for thought

Food for thought

Thanks for the Master Class Dan. Your post gave me a Mark Shubin flashback.

I had never really thought through the differences between causing silver halide crystals on emulsion layers to react vs lighting up photosites on a CMOS sensor from a purely optical perspective. Moreover, your comments on the inter-relationship between the resolving power of a lens, the CofC and the best possible level of OLP filtering for a particular format really expanded my understanding.

I guess the proverb is true; the more you know, the more you realize how little you know. :matrix:

 

[Barry Green]

Graeme, you might want to have a chat with Adam Wilt; his measured resolution with a Zeiss SuperSpeed is 2.5k, whereas you're observing 3.2k. That's a pretty big gulf, and it'd be nice to know where the disconnect is.

 

[Graeme Nattress]

Looking at Adam's charts, and I don't know for certain, but they look as if shot pretty wide open, which makes both lens and focus the most likely limiting factor. I think Jarred shot the charts I was using, which are 4k charts, so no futzing with moving in and out, and they contain both 4k and 2k linear sinusoidal zone plates, for very easy and accurate resolution measurement. Basically, you just plot a scanline in imageJ and see where your modulation drops off, and read off the linear distance at which that happens, divide by the total radius of the zone, multiply by 4k or 2k and there you have it! It was a Cooke 65mm lens if I remember rightly, that was used.

2.5k / 4k is 62% which is very low for a bayer sensor. Usually you're looking at 70% + (which is what I said very early on we'd get). Depending upon demosaic algorithm, I've measured anywhere from 72% to 78%, and actually beyond (at the expense of some artifacting I didn't like, so I backed it off at bit to live at 78%, which with the current scheme was the best I could get without getting that pixel level artifacting I don't like.)

I'm pleased Adam saw the chroma denoise / moire destroyer. I think that's a pretty key thing to have with any bayer sensor, and what we have in there works really well, even though I say it myself.

It looks however, that the highlight control in Redcine wasn't working right for some reason in his system, on the colour shift on the over-exposed regions. To debug that, I'd need an R3D file though....

Graeme

 

[Kenn Christenson]

A little disconcerting that 2K only comes in around 720P resolution. Perhaps it is also an issue with the lens

 

[Miltos Pilalitos]

It looks however, that the highlight control in Redcine wasn't working right for some reason in his system, on the colour shift on the over-exposed regions. To debug that, I'd need an R3D file though....
Graeme

The Highlight control doesn't work in any one's Redcine that i know of. I am talking about PC land. It used to work in a very early Redcine built but not in the current PC version.

 

[Graeme Nattress]

That could explain a few things then....... I'll look into that. Thanks Miltos.

For 2k, I'd expect about 1.6k which is significantly more than 720p or 1080i.

Graeme

 

[Miguel "Macgregor" De Olaso]

Grame said: "I'm pleased Adam saw the chroma denoise / moire destroyer. I think that's a pretty key thing to have with any bayer sensor, and what we have in there works really well, even though I say it myself."

http://provideocoalition.com/index.php/awilt/story/three_three_letter_cameras_ex1_f23_red/P2/

RED ONE chart doesn´t look good because of the color aliasing, even when 4K was resampled to 1080. Why is this happening? Why does the F23 resolves a lot more detail (not talking about sharpening)?

 

[Deanan]

There's alot of details missing about the tests and without actually being present for a test and the post processing, I wouldn't trust any test I wasn't at.

In any test there are all sorts of biases based on past experience guiding judgements or anomalies in processing. Adam's results are very far away from the tests we've done. Without being there, we can only guess. This kind of comment hints at the conditions of the test:

So our tests are fragmentary, the cameras weren’t fully set up, the charts were sloppily lit, and there are some mismatches in focal lengths and framing.

As Sean Fairburn would say, test it yourself...

 

[reality]

So our tests are fragmentary, the cameras weren’t fully set up, the charts were sloppily lit, and there are some mismatches in focal lengths and framing.

Wow. All I can say is that's quite irresponsible for a review to be that lazy especially since Adam is quite well respected. What's the point of a long detailed review if it's going to be incompetent. At least take the time to learn the f23 and red.