David,
Can the concept of pixel skipping be viewed similar to how AVCHD compression was viewed a few years back? I recall reading posts across many forums complaining about how there was no future in using AVCHD... now its use is widespread. Would it be accurate to say that many next generation cameras -- perhaps just prosumer at first -- will likely be using some means of assimilating MOST but not all of the pixels that the camera sensor is reading? Could there be better implementations of pixel skipping just as there are now better implementations of AVCHD compression than there were a year or two ago?

It seems like many people are currently wondering if the AF100 is something to jump at or does one of the major camera companies have what you're referring to just around the corner. It certainly always comes back to needing to make the decision to "pull the trigger" at some point and just purchase the best technology that is available at the time. I just off-loaded all my HDV equipment, and my plan is now to use the GH2 I just purchased to work as a pair with the AF100. But if the AF100 is using a stopgap measure that will be addressed in 6 months by a different camera, I'll wait. It'll be interesting to see if the AF100 (or the Sony you're referring to, or perhaps the next Canon??) has the answer to what you're asking:

"Can it be categorically stated that the AF101 reads out all 12.4 million photosites each frame, forms the image from all of them, then downresolves the result to form the final 1080 image"?

I guess a follow-up question is the one you also ask above: Does it matter though... if you have a fast lens and if the video you're capturing is great footage? Would the issue be that there would be minimal professional use for the footage due to the limitations you're mentioning?

James - what I think is important to realise is that pixel-skipping has got a bad name because it's associated with the moire/aliasing in DSLRs, not because it's inherently a bad thing. But that's due to their lack of an OLPF - not a problem in a made-for-video camera such as the AF100.

In picture quality terms it's something of an irrelevance - as long as the right OLPF is used. That's why nobody can look at a camera output, see no moire/aliasing etc, and draw the conclusion of no pixel skipping.

The *GOOD* thing about pixel skipping is that it enables a made-for-still sensor to be fairly easily adapted for video use at quite lost cost and low power consumption/heat generation. The *BAD* thing is regarding sensitivity issues - which can manifest as either/or signal-noise or ISO rating.

In terms of the future, I'd expect the way forward to be more in terms of made-for-video sensors with an appropriate number of photosites for the output resolution (more like 3-4 megapixel than 12), rather than improving the way what are fundamentally still camera sensors are adapted.

A month on, I think it's now possible to be a lot more certain about what the AF100 does and does not do in a technical sense.

Firstly, I believe it is correct to say it is NOT line skipping. I now believe that what it IS doing is "pixel binning", as opposed to "pixel skipping", That should give better sensitivity results, but worse resolution. That means it is reading the entire sensor, but not forming a high res image and downscaling - "pixel binning" processing is far simpler. (That's why the cost, power consumption, etc are what they are.)

It also seems to not have 12.4 million photosites making up the active frame, but rather 13,932,800 in 16:9 framing, so nearer 14 megapixel. That works out to be 4976 x 2800, and this figure can be easily obtained from the GH2 specs. (The 12.4 million would be the figure if a GH1 chip was used.)

Put it on a test chart, and the measured resolution comes in far less than I originally expected, roughly equivalent to a 720 sensor. This is in line with various posted images which seem to show it looking somewhat soft compared to cameras with 3 1920x1080 sensors. Looking at the null points of the aliasing, it seems to give way to aliasing at 1244 pixels (622 line pairs) horizontally, and 700 pixels (350 lp) vertically. The chart is symmetrical - so no line skipping.

The eagle eyed will have spotted that that resolution is exactly one quarter in each direction of the sensor pixel count. The most likely deduction therefore must be that it's pixel binning on the basis of 4x4 blocks of photosites. Hence 8 green, 4 red and 4 blue photosites in such blocks are having their charges collected (binned) into 3 "bins", one each for R,G,B, before these 3 values are digitised and processed.

Hence, the image seems effectively formed from 1244x700 "super-pixels" each with an R,G,B value, and each the averaged result of 16 sensor photosites. It's therefore equivalent to a 3 chip design with 3 chips each of 1244x720 or about 0.87 megapixel each.

Gentlemen,

The camera does not do pixel binning or line skipping. It uses an optical low pass filter to resolve the moire and aliasing.

Thanks,

Jan

Thanks Jan, however this does not explain many of the results being obtained in tests. For example, if the camera is not pixel binning, this doesn't explain why the camera is only really producing around 650-700 TVL of good resolution, rather than 900-1000 odd TVL that one would expect from a true 1080p camera.

Would it be possible for you to tell us or find out from the engineers how the data is being read out and interpreted?

Well every AF100 I have seen up on a res chart is more like 800 lines. And no I am not in a position to tell you proprietary engineering information. I just know that it does not pixel bin, nor line skip.

Thanks,

Jan

Hi Jan, here are some extracts from a chart. I realise lens choice has an effect, but information I have suggests that the results shown in these chart extracts are being replicated. I would be interested as to why, for example, the chart shows some very strange issues, such as why there is still aliasing showing even when there is no apparent detail shown (eg in the circular section of the chart I have attached)?

This cannot be put down to a lens or focus issue because for aliasing to occur, there has to be detail in the first place.

How were these stills produced? They are very small, 227x354, not HD resolution, therefore unreliable examples of the problem. Could you post 1920x1080 stills of the charts?

They are 100% crops of a res chart. Totally reliable I assure you. Though they had to be recompressed for the web, the detail on show is the same as the original. As I mentioned, others are finding similar results.

I didn't post the whole chart because it isn't necessary. The crops I have posted are the focal points of the issues I mention.

Jan, the problem with standard resolution charts (with just horizontal and vertical lines) is that it's extremely difficult to tell the difference between real detail and aliasing. I assume those are the res charts you've seen? You know the input to the system is 800lp, you can see an output for that block, you therefore think "oh, it's resolving 800lp". This is the case with the second image Simon Wyndham has just posted.

Use a zone plate (or, like Simon Wyndhams examples, at least a circular resolution chart) and it all becomes vastly clearer. True (or "real") detail resolves as it should according to the original chart, aliasing shows as a false,arc whose centre appears to be somewhere other than the centre of the pattern. Simon Wyndhams first image shows the aliasing well. On the fourth ring (the 800lph ring), the circles appear to have a centre to the left of the chart - they are clearly aliases.

Use a true zone plate (rather than a circular resolution chart with bands) and it becomes far easier to do measurements - all Simons chart tells you is that it can resolve 600lp, but it can't resolve 800lp. As I said in the previous post, it seems to be very close to 625 line pairs (Simons previous estimate of "650-700 TVL of good resolution" is actually over optimistic!). If I could measure it accurately enough, I'd put a very, very large bet that I'd find it to be actually 622 lp, as that correlates with exactly one quarter of the sensor pixel count.
If they are 1:1 pixel crops, (rather than downscales) they are fine. They agree in essence with similar examples I've seen, and other people have as well.

Things to bear in mind.

Lens matters tremendously, you also have to be really sure of the focus. Not always easy. Slightly OOF and result is meaningless. Sharpening can cause problems and viewing on a computer screen can cause moire.

I will run some checks with charts next week using different lenses, and I guarantee I can get different results. I will post my findings here.

I will also compare HDSDI out direct recorded to NF and to AF100 native CF card.

You also have to try different scene settings to be fair.

Not true. In the case of a too soft lens, or being out of focus, all that happens is the amount of aliasing drops off, as does the mtf of the coarser (true) detail.

What doesn't change is the aliasing pattern in terms of fundamental shape, and where the aliases appear to be focussed. (The centre of the circle of which the arc is a part.) That's fundamental to the sensor pattern, and how the image is derived.

I believe this is exactly what Simon meant by "This cannot be put down to a lens or focus issue because for aliasing to occur, there has to be detail in the first place."

I don't see how scene settings are likely to affect max resolution.....?

So did you shoot this, and what chart was used?
Or is this someone else's test?

Thank you for clarifying that.

These images are in focus otherwise the aliasing pattern would not be there. It doesn't seem to be a lens issue. How do the detail adjustments affect this aliasing issue? Is the low pass filter adjustable?

Simon Wyndham: tests demonstrate "camera is only really producing 650-700 TVL of good resolution, rather than 900-1000 odd TVL that one would expect from a true 1080p camera,"

Jan then replied that she sees Res charts with more than 800 lines.

I'm sure Jan is being forthright. Are there circumstances where these different results could occur?

My main question: for a point of reference, what would a 1080p camera in the same price range -- the EX1R for example -- have for lines?

If I remember correctly (I have an EX3), about 1000.

Yep, the EX series is just shy of 1000, and with minimal aliasing.

The apparent contradiction is easily explained. Look at Simon Wyndhams second photograph - the one with the resolution blocks of horizontal lines. You can clearly see lines seemingly "resolved" on the 800 line resolution block. I'm sure this is what Jan is seeing.

Trouble is - it's aliasing. It's "false" - apparent - resolution.

That's obvious from Simons other photograph and the circular rings, especially the ring corresponding to 800. When near horizontal, the lines are bent the opposite way to the lower rings for resolution - their apparent centre is not the centre of the chart. They are aliases. They don't truly represent the pattern on the chart.

There is also another way to think about it. Look again at Simons second photograph. ( af101-horizontal.jpg ) In the blocks 200, to 400, to 600, the lines get closer together - as they should, same as the real chart. But look closely at the "800" block. The lines appear to not be as close together as in the previous (600) block. Why? Because they are aliases.

You can just still see "detail" in the 1000 block - but the lines are then even further spaced apart. Exactly what would be expected from aliases.

For this camera, it can resolve 600 TVL - it can't resolve 800. It appears to - but is just aliasing.

Stupid question (as I'm deciding between the EX1R and the AF100): for approximately the same price, I could buy an AF100 along with a NanoFlash unit as compared to an EX1R. Wouldn't using the NanoFlash with the AF100 and being able to record at such a high bitrate with 4:2:2 color raise the quality of the AF100 video capture above what the EX1R captures natively? (I know it's a newbie question... I'm just not sure what happens when in the capture process to capture higher resolution).

The Nanoflash can only record what the camera gives it. So if the AF-100 captures 700 or 800 lines then the Nanoflash will record 700 or 800 lines.

What the Nanoflash does for you is record those 700 or 800 lines with more information which leads to a closer representation to the uncompressed signal.

Actually I am seeing resolution in the DSC chart and it is not aliasing. You cannot be sure about what I am seeing as you are not here in my lab looking at my setup and my DSC chart.

Thanks,

Jan

Where can you purchase an Af100 and nanoflash for less than a exlr?

Well - can I ask how you therefore explain Simons chart that we are all able to see?

Which clearly DOES seem to show the 800 line result to be aliasing? And which seems to correlate with other results that have been seen?

Is it possible for you to post an image of a test chart (ideally a zone plate) showing exactly the results you are obtaining?

Not sure what Simon posted, didn't look at it. Keep in mind that resolution is a factor of more than just the camera. You need a lens that will resolve the resolution as well. We are using a backlit DSC chart, a Zeiss 50mm Compact Prime and the AF100 set at 400 ISO.

Best,

Jan

Jan, you really should look at the images I posted (and it is, if you don't mind me saying, a little curt to dismiss them out of hand and hold a discussion on the matter without even looking at them). Aliasing is visible where there is no apparent detail at all, which completely rules out the lens or focus. The lens was clearly resolving the detail, but the camera was reducing it all to aliasing.

Further, as David mentioned, others who have been performing tests have been getting similar results. I do not think there is any need to be defensive about this. Science is about actual results, not emotional disagreement. Without seeing your results, especially a zone plate which would prove what you are saying outright, we cannot take what you say as read, especially when the evidence that is on show says otherwise. Incidentally, the chart used in the shots I posted was a front lit DSC chart. Other settings are irrelevant in this case AFAIAC. If the lens was firing an even sharper image into the camera the aliasing wouldn't go away. Some reduction in aliasing might be achieved by lowering the enhancement in camera, but as you can see there is aliasing even on areas where detail has been reduced to an almost flat grey fill.

There has been absolutely no "emotional disagreement" anywhere in this discussion, but
it's just that type of baseless accusation that results in closed threads and locked accounts.

Don't do it again.

So let me ask the question one more time Simon, did you shoot the chart or was it passed on to you?

Our results were as Jan had mentioned. We tested the AF100 on several charts, in several scenarios, and under varying conditions. It passed our tests with flying colors and we put in into use with our Red.

You may be into the "science" of camera specs, which is all well and good.
But we are into making money in the commercial business. And this camera paid for itself in under a week! IMHO, this camera is a no-brainer, no matter how many lines of resolution "you" are seeing on a chart.

All the Best!

Dave

Hi Simon,

So which lens did you use? I think it is pretty odd as I know that my engineer , Barry Green and I all saw 800 lines. Barry Russo, said he would set it up again and do a frame grab for me, but it might take a couple of days. Settings were right out of the box without tweaking. I thik the lens might be a key influencer on this though.

Best,

Jan

I really think it would be a good idea if you did, Jan. The results are pretty conclusive, especially regarding the alias patterns from the circular resolution rings.

Did you read what I previously put? (Post no 50.) If not:
Again, this has been covered before. (See post 52.) A different lens will give a different mtf to the result - but will do NOTHING to affect the shape of the alias patterns, if any can be seen. Their shape is solely a function of the native camera resolution.

And as Simon says, the mere fact that alias patterns CAN be seen must indicate that it's the native camera resolution that is the weak link - NOT lens resolution.

If you dispute the results that Simon posted, if you think he has done something wrong, then please, would you like to show us what you have seen?

[EDIT - Having just seen your last answer, can I further ask if your chart had circular rings of resolution (as Simons example) or just vertical and horizontal lines? The former shows up what is true detail and what is aliasing easily, the latter doesn't do so anything like as well. If it's all you've got, try panning the camera slowly from side to side. If the 800 block lines appear to ripple in the opposite direction to the direction the block is moving, they're aliases.]

Okay, So I took a look at the charts that Simon posted. And as soon as Barry pulls it together I will post our results but they do not look like Simon's results.

Best,

Jan

Thanks Jan - I look forward to seeing them.

After the last paragraph of my last post, it occurs to me that if you only have a chart with horizontal and vertical res lines, then could you also post a couple of seconds of movement - very slowly panning across the chart? So we can be unequivocally sure whether we're seeing aliasing or real detail? Obviously, a chart with circular resolution or a zone plate would be far better.

The effect ("If the 800 block lines appear to ripple in the opposite direction to the direction the block is moving, they're aliases") is analogous to the wagon wheels on the stagecoach appearing to turn backwards. (That's also aliasing, though wrt time, not space, as in this case.)

Quote from David Jones in post #67:

"Our results were as Jan had mentioned. We tested the AF100 on several charts, in several scenarios, and under varying conditions."

David: are you also able to post your results?

We burned that footage long ago. No need to keep it around as we have put the camera into production work. We captured to SDXC card as well as SDI into a Kona card and determined that for the majority of the stuff we would shoot with the AF100, SDXC would be fine. For higher end work & keying we can use the Red. As far as footage goes, we do broadcast work which we never post on places like youtube or vimeo.

By the way... This thread has gone way off course.
Original question... Does the AF100 line skip?
Answered by Panasonic Rep... No.

All the Best!

Dave

I disagree that the thread has gone off course. It's possible to infer quite a lot from properly done test charts about how a camera is working internally, how it's reading the chip. If we accept Simons results as valid, then one thing they conclusively prove is that it is *NOT* line skipping - as I said many posts ago, see post no 40.

Line skipping would not give a symmetrical hor/vert result to zone plates or Simons chart. Of course, the obvious question then is "if not line skipping, what is it doing?"! That's the current validity to the thread.

The most likely theories I've had presented to me revolve around some form of pixel binning (which the GH2 - same sensor - certainly DOES do in some modes) - but Jan denies that. So if not that, just what is it doing?

I think the next step is to see whether any further tests show different results to Simons, or tend to substantiate what he has posted. Let's wait and see what Jan comes up with - and may I ask how Olofs tests with an AF101 and charts are going? (Post 51.) Do you have anything yet to share with us, Olof? (And do you have charts with circular resolution rings?)

I am enjoying this discussion very much. A lot of important technical issues are being discussed here. Regardless of the resolution issues, the camera puts out a very, very impressive image.

Yeah, this reminds me of the big arguments regarding the Canon 5D and other DSLRs. People saying you can't shoot this or that with that camera. Two years later its all being done by professionals--- professionals that are able to take what the gear has to offer and use it as a tool, and gets some beautiful images out of it and into commercially accepted projects.

There is no doubt that this camera is intended for the indi film maker and small studio. Panasonic has always catered to this crowd. While I am not sure if this is my future camera, I think Panasonic deserves major kudos for pushing this 4/3's format and for pushing the market forward into the large chip and interchangeable lens era. Sony (already moving that way) and Canon will have to follow in their video lines.

"I am enjoying this discussion very much." Agreed. Even though there's a spirited back and forth, I think the resolution (no pun intended) of the questions in this post are very important. I personally only purchase a camera at this price range only every 3-4 years, so it's important to determine if the resolution is as high quality as possible to open up as many distribution options as possible.

In the long run, there would be no upside to Panasonic indicating all over the internet and their own documentation that the resolution is 800 lines if it wasn't so, as it would certainly be discovered at some point. But I guess the proof will be in the pudding, as they say, and it'll be interesting to see the charts as described.

I've had the benefit of reading this entire thread from start to finish in one go.

Simon said posts and posts ago that the image he produced was a 100% 1:1 crop of the full 1920x1080 image, not downscaled.

I ask, can we see the entire 1920x1080 image, not a crop or downsampled for the web? That wouldn't be a heck of a lot of trouble, would it? Considering there is so much debate here on that very image, I see that it's only fair we see the entire frame and details about what lens was used to reproduce it. The info we have on it thus far is minimal.

It's hard to know what to make of the the AF100 sensor. It's pretty amazing that no one at Panasonic knows or is able to tell us how the AF100 sensor works. It's visibly similar in size to the GH2 sensor with same size pixels and and released around the same time, and Panasonic says it is not the same sensor but cannot tell us even one way in which it is different besides possibly OLPF, which isnt really the sensor itself. Why put a 12+MP sensor in a 2MP camera unless you already had a 12+MP sensor sitting around?

Panasonic says the AF100 does not line skip and does not pixel bin? Is there some other option besides reading all 12MP at 60fps? it seems hard to imagine, though maybe not impossible i suppose, that the camera is processing data from the sensor at >1GB/s without extensive cooling or power consumption. Panasonic answering the line skipping question so decisively but with no explanation given the circumstances seems to just bring up more questions. Additionally, I am not sure I understand why answers about line skipping always are followed by information about an OLPF. Jan makes it sound like there is a special OLPF that helps avoid aliasing, but if there is no skipping or binning, it should not need a super special OLPF. Just an light OLPF designed for the full 12MP sensor should suffice, since it is line skipping that would cause a need for an extra strong OLPF to avoid aliasing in the first place.

Maybe they just mean "No, you wont see artifacts of line skipping as strongly as you are used to with other line skipping cameras like the 5D/7D" or "No, we do not skip lines, but we do bin them which causes similar artifacts but helps with noise, but the aliasing aftifacts are handled by a strong OLPF".

I hope they mean "No, the sensor reads all 12MP at 14bits and 60fps at 720MP/s and processes the data at full resolution at 1.2GB/s and does a high quality downscale to 1080p for encoding and recording, and all that stuff we said about the OLPF actually isnt that important because the images are so oversampled that there is incredible sharpness with no aliasing anyway." It seems unlikely, though, especially after seeing it moire a bit on charts.

In the end, I guess what matters is that we know the practical realities of the camera: It does show some aliasing but not too much unless you shoot charts and crank up detail. It's not magic, its performance in terms of latitude and noise are, like most of its competitors, proportional to its sensor size and retail price.

It doesn't matter that we dont really know how it performs as it does, but it is a little bothersome that the company that designed the camera wont tell us some of its basic design features.

Hi,

The Zacuto Shoot out is in process, and frankly by the time I get to this little task that may well be published. I know it will be shown at NAB.

Please know that I have a very busy job that requires a lot of extra time and having these chart thrown into it is just not on the schedule as there are other things that I am having to do that are more time sensitive. I simply just don't seem to find the time to set all of this stuff up and do this. Sorry.

Knowing that all will be proven with the Zacuto Camera shoot out, is probably the best thing. So with this, I bid this list adieu. See you all around the web.

Thanks,

Jan