Aliasing and moire

[Alister Chapman]

I've been testing and playing with my F3 and first off let me say this... I love this camera, it produces amazing images and I can play with lots of lenses!

But, it's not all roses. The F3 suffers from aliasing. A zone plate shows extensive aliasing. Until today this had not caused any concerns as I had seen little evidence of it in actual footage, but when shooting some brick houses this afternoon I came across some coloured moire patterns appearing as faint coloured stripes across the brickwork in the footage. It's not anywhere near as bad as a DSLR, my wife looked at the footage and didn't notice it until I pointed it out to her, but it's certainly there. This is disappointing on a camera at this price level, my EX1 doesn't do this. Now the zone plate shows this to be an issue with the cut off of the optical low pass filter, so I doubt that there is much that can be done in the firmware, but then Sony have done some clever stuff in the past with firmware updates. When working on my picture profile settings I did find that increasing detail above -15 would increase the visibility of the aliases on the zone plate, however when I tested various detail settings with the brickwork there was little difference. I think Nigel Cooper has also seen this, but I've not seen it mentioned anywhere else. Has anyone else observed this?

[Timur Civan]

yea i ve seen it.

F3 Resolution w/ Cooke 50mm @ T5.6 on Vimeo

The very smallest circles have blue aliasing. IT happens, but its not awful. So long as straight lines dont exhibit stair stepping, and the actors hair line isnt twittering in a medium close up im fine with it.

[Alister Chapman]

I've attached a grab from a rather under exposed zone plate which shows the multiple alias frequences and a crop from a frame of the houses. I've circled the colored moire which manifests itself as purple and red stripes on the bricks which move with any camera movement. They are not particularly bad but, they shouldn't really be there at all and spoil an otherwise very clean image.

[Timur Civan]

I would HARDLY call that extensive aliasing.

Every camera will alias. Thats the point of the Zone chart. Its to determine, When not if.

essentially in a digital sensor its impossible to prevent aliasing, unless you are softening the image so much with an OLPF that its rendered soft. The AF100 comes very close to this threshold. The OLPF eliminates "aliasing" but leaves the Af100 a touch soft, but at the limit of pleasing. Sony has always erred on the side of maximum resolution, and thus you get some minor artifacting.

[Jim Tittle]

Makes the building look better.

[Daniel Doherty]

Alister,

Are your tests using the onboard 35mb/s codec?

Do you think recording to an external recorder with higher bitrates would provide better results?

Thanks,
Daniel

[Alister Chapman]

It's not the codec. The aliasing is happening at the sensor.

The F3 exhibits a lot of spacial aliasing. I'll set up some zone plate properly tomorrow and take a closer look, but the indications are that there is either no OLPF or the cut-off is too high. The moire on the buildings is more noticeable on a moving shot as it rolls up and down the buildings. Such artefacts are absent from cameras like the EX1 and EX3 which have correctly designed bi-refrigent filters, yet these are half the price of the F3.

A correctly designed camera should not alias, it's only when compromises in the design are made that aliasing occurs, normally as a result of trying to squeeze more resolution out of the sensor. While it is not uncommon for such trade offs to be made, it is not normally so bad that just shooting the front of a building leads to moire. Imagine what a striped shirt, tie, or patterned dress might do.

[Timur Civan]

Testing time!!!!!!!!! :) my 3rd favorite thing!!!!

[David Heath]

Originally Posted by Alister Chapman

I've attached a grab from a rather under exposed zone plate which shows the multiple alias frequences .........

The only aliases I can see there are on the diagonal, and centred around the highest frequency. Frankly, I was expecting far, far worse after reading your first post.......

Being on the diagonal, I suspect they may have a lot to do with the Bayer patterning, and as such are little more than may be expected. It may get better with a higher pixel count chip - but likely at the expense of sensitivity, price, complexity etc.

Questions I would ask is why an underexposed image? (Which also looks a bit noisy.) And whilst the grab seems to be full raster 1080 still image, there are no units on the zone plate - which means it's not really possible to definitively say what spatial frequencies the circles relate to. Or can you assure us that it is set up such that the outer rings correspond to 1080 lines? Finally, is it a full colour image, or just luminance?

Originally Posted by Alister Chapman

The moire on the buildings is more noticeable on a moving shot as it rolls up and down the buildings.

What's key about an alias is that as you pan the camera, the alias will move the opposite way to what you would expect, and opposite to the way the subject is moving. That's the easy way to tell aliasing apart from "real" detail if you only have horizontal and/or vertical lines to look at. (And why zone plates make camera testing so much easier.)

This fact about aliases moving the "wrong way" is also why they are so undesirable - they confuse the motion sensitive aspects of a codec, the codec sees an object moving both ways at the same time. Result - confusion. (And wasted bitrate.)

[EDIT - Just a word of caution, and that's for anyone not to get confused with genuine aliasing of the sensor, and aliasing of an interaction between the image and a computer display! It's easy to tell the difference - just drag the image size and see what happens. If the alias remains constant, it's a real one, if it changes pattern, it's a computer display issue!

Best way is to import Alisters image into Photoshop and set a 1:1 pixel display mode. Doing that the most obvious (real) aliases are on the diagonal, though there are a couple of very small ones on the vertical I now see. To be a complete test, you really need to present it with a range of frequencies BEYOND the system resolution. Ideally, the result will be plain grey. Any patterning or lines must be aliasing. I suspect that in Alisters image the highest frequency may not go as high as 1920 lines horizontally, let alone above?]

[Alister Chapman]

David, the zone plate I posted was just one I was able to grab quickly. It does not go out past the extinction point so is not showing the full story. I'll post a correctly exposed and scaled example this morning, although I don't have a calibrated set so I'll do a range of sizes so we go well past 1920. When I was looking at the detail and aperture settings earlier in the week I observed strong aliasing beyond 1920. What the underexposed plate shows is that even when you are well below maximum contrast aliases are still visible at lower resolutions than the system resolution which is not a good sign.

Aliasing as you say plays havoc with transmission codecs.

[Peter Moretti]

Originally Posted by Alister Chapman

...
A correctly designed camera should not alias, it's only when compromises in the design are made that aliasing occurs, normally as a result of trying to squeeze more resolution out of the sensor. While it is not uncommon for such trade offs to be made, it is not normally so bad that just shooting the front of a building leads to moire. Imagine what a striped shirt, tie, or patterned dress might do.

Alister,

The EX's do alias. Look at page 14:

http://thebrownings.name/WHP034/pdf/...X1_and_EX3.pdf

That said, I applaud your willingness to call a spade a spade. And from what you're saying, the F3 aliases more than EX's. This could be a real problem... or not so much of one. I hope Sony does do something about it. As the F3 images I've seen are simply lovely.

[Alister Chapman]

OK, The EX1 does alias a little, but it doesn't cause any problems. The attached image is a montage of Zone Plates at different relative sizes (changed by zooming in/out) for both my PMW-F3 (Fitted with a Nikon 18-135mm DX lens) and a standard EX1R. It's not totally scientific, the zone plate is a home made print done with an inkjet printer, but there are no printer artefacts visible on the printed sheet.

All the frame grabs are 1:1, no scaling or adjustments.

To me it clearly shows a lot more chroma aliasing from the F3 both horizontally and vertically as well as on the diagonals. In many respects this is expected to some degree, after all this is a bayer type sensor and the usual trade off in cut-off for the optical LP filter applies, trading luma resolution agains chroma aliasing, but frankly I'm disappointed by what I'm seeing. Luma aliasing appears comparable between the two cameras.

Anyone looking at these does need to look at them 1:1 pixels or zoomed in to be sure that they are not seen additional aliasing from the computer screen.

[Peter Moretti]

Alister,

Thanks for the pics. If this is mostly a chroma issue, then it's also seems possible that 4:2:2 or 4:4:4 might help to aliveate it.

But the F3 has a relatively low pixel count sensor. And pixel count is pretty inversely related to resolution. And resolution is pretty inversely related to aliasing. All that means is we are fools to expect too much from the F3 resolution wise. Just my two cents/pence ;).

[Alister Chapman]

4:2:2 and 4:4:4 won't make any difference as this is in effect occurring at the sensor. It is possible that Sony could improve or adjust the de-bayer algorithm to try to help, but it's mainly down to the cut off of the OLPF being too high.

Yes more pixels would probably have helped. It's a shame because in almost every other respect I really like this camera. It's not going back, I'm keeping it, but I do need to keep this in mind when shooting with it. Good on set monitoring will be essential.

The F35 aliases a lot as well, but also produces great images.

[David Heath]

Originally Posted by Peter Moretti

If this is mostly a chroma issue, then it's also seems possible that 4:2:2 or 4:4:4 might help to aliveate it.

No. Aliasing is an artifact that gets burnt in right at the sensor of the camera. It's where high frequencies get folded back to appear like lower ones - hence the very name "alias". Once that happens, they are of a frequency low enough to be little affected by such things as colour sampling etc. Anything you do beyond the sensor will affect the "real" picture more than the aliasing.

Similarly, using different lenses, even defocussing, WILL *NOT* affect the alias PATTERN, though they will affect the intensity of it.

It might help to think of the audio equivalent. If a system is sampling at 20,000 times a second, it sets a limit of 10kHz as the absolute upper limit that it can resolve. This is the Nyquist Limit. If you present it with a frequency higher than that, it will alias, in a way that folds back around the Nyquist limit.

So feed in 12 kHz, and the output will be 8kHz. Feed in 18kHz, and you hear 2kHz as an output. Maybe most interestingly, if you slowly increase that frequency, the output frequency will slowly DECREASE correspondingly! Take it a stage further, and at exactly twice the Nyquist frequency, you get 0Hz, or D.C.

Moving back to zone plates, the equivalent is the centre of the alias circle, and on calibrated charts that lets you see where the resolution corresponds to twice the Nyquist rate - or twice the effective sensor resolution.

Once you see circles centred on the diagonal, it's sign that it's not quite such as simple model as such as an EX sensor, and the same goes for coloured aliases. In the case of the F3, I'd take it as primarily caused by the Bayer patterning on the sensor.

Originally Posted by Peter Moretti

But the F3 has a relatively low pxel count sensor. And pixel count is pretty inversely related to resolution. And resolution is pretty inversely related to aliasing.

True to a point, but you have to take in to account how the sensor is being read. That statement is only true on a like for like basis. At present, it's not really viable to read a high pixel count sensor at video frame rates without it costing a fortune. Hence manufacturers "cheat", and read the sensors in such a way that they seem to have a lower count than they really have.

Zone plates effectively give away some of the secrets that manufacturers wouldn't advertise. They reveal the real EFFECTIVE resolution of the sensor, and that may be typically a quarter of the pixel count, for very good engineering reasons.

If you don't believe me, try pointing any video DSLR, or an AF101, at a zone plate........