stuttering or aliasing on hard edges

[Bryan Harvey]

Im not sure what the proper terminology is, but in my testing of my brand new 5d2 I see serious stuttering on sharp edges (like the horizontal edge of a table), anything with a straight line. Most noticeable with camera movement.

specs:
5d2
any ASA
1/60th shutter
sandisk III 16gb card
A/V port mini plug to Panasonic 9" monitor.

This is visible during recording and playback as well.
Is this a faulty camera or is there a possible camera setting I have overlooked?

Thanks!

Bryan Harvey

BryanHarveyFilms.com

[Douglas Joseph]

On prolost there's an article about this issue, and how to kind of resolve it.

ProLost - ProLost Blog

[Daniel Browning]

Originally Posted by Bryan Harvey

Is this a faulty camera or is there a possible camera setting I have overlooked?

Not a faulty camera, nor can it be fixed with settings. The video on a DSLR is generated by skipping 2 out of 3 rows, which causes very terrible vertical aliasing: far worse than any camera I've ever seen.

You can try to blur the image to reduce aliasing in post, but that may not be enough (especially for chroma aliasing).

The only real solution is to blur the image optically before recording. Even if you smear it enough that the effective resolution is less than Standard Definition (480p), there will still be some aliasing, but it will be down to acceptable levels.

[Charles W. Hull]

Originally Posted by Daniel Browning

Not a faulty camera, nor can it be fixed with settings. The video on a DSLR is generated by skipping 2 out of 3 rows, which causes very terrible vertical aliasing: far worse than any camera I've ever seen.

First, I agree the 5DII has inadequate aliasing performance. But I question whether the explanation is as simple as skipping rows. Video is inherently aliasing - it is up to camera maker to implement adequate anti-aliasing algorithms to minimize this, and this gets down to how much processing is devoted to this function. Obviously the 5DII has anti-aliasing, it's just not good enough. Anti-aliasing algorithms are designed to trick the eye by bleeding in intermediate brightness pixels at the right places on sloping edges. There is no way of knowing whether or not Canon uses information from all the available pixels (rows and columns) in their calculation (unless they said this and I missed it). Even if they don't it would be possible to have good aliasing performance, given enough processing.

I was hopeful that the 7D would perform better with its two processors, that some of the added processing would be used to beef up anti-aliasing, but it doesn't look like it. 7D video looks pretty much like 5DII video. I expect it is a thing where if enough of us complain Canon will reduce aliasing in future cameras. In the mean time I notice my audience isn't complaining about the aliasing. (But it bugs me!)

[Daniel Browning]

Originally Posted by Charles W. Hull

Video is inherently aliasing

That's one way to put it. If the video doesn't use protection, it might get aliasing. :) Video does not alias at all when sampling is at a higher rate than the spatial frequency of the analog scene detail. Whether or not that will be the case depends on circumstances like subject, focus, motion blur, diffraction, abberation, and optical filtration (e.g. low pass AA filter).

Originally Posted by Charles W. Hull

- it is up to camera maker to implement adequate anti-aliasing algorithms to minimize this, and this gets down to how much processing is devoted to this function.

There is no such thing as an anti-aliasing algorithm. It's not even *theoretically* possible. Software cannot know whether the red and green streaks in blond hair are from the art department or from aliasing. The very best software I've seen, Capture One Pro, works on the raw data and is slow as molasses (think seconds per frame, not frames per second). But it just smears the detail into blobs. Smears and blobs may be an improvement over sparkles, but it's still nowhere near as good as optical filtration.

Originally Posted by Charles W. Hull

Obviously the 5DII has anti-aliasing, it's just not good enough.

Well, "good enough" is the in the eye of the beholder. Many video shooters *love* aliasing artifacts, so this aspect of the camera is a big feature for them. Personally, I abhor it, but I still shoot the camera anyway because I know that much of my audience would prefer it over the alternative (XH-A1).

Originally Posted by Charles W. Hull

Anti-aliasing algorithms are designed to trick the eye by bleeding in intermediate brightness pixels at the right places on sloping edges.

Sloping edges are just one of many ways that aliasing is visible.

Originally Posted by Charles W. Hull

There is no way of knowing whether or not Canon uses information from all the available pixels (rows and columns) in their calculation

There are several ways to measure it, in fact. For one, you can compare the noise of the video image with a downsampled still image. If all the pixels were used, the random noises would add in quadrature and SNR would increase by a factor of 2 for each four-fold decrease in megapixels. The fact that video has far more noise than a downsampled still indicates row skipping. It would be important to use a photon shot noise dominated image tone to avoid the possibility of read noise differences in on-chip binning.

Second, the zone plate clearly shows far more aliasing vertically than horizontally. Compare that to the results you get from typical Bayer on-chip binning result and you see something completely different. Not only is the aliasing more muted in on-chip binning, but it's equal in both directions.

Third, you can see what kinds of aliasing artifacts result. Ringing, softness, aliasing, sharpness, etc. can point to the type of resampling filter used. Demosaic artifacts (e.g. directionality errors and mazing) offer clues as well.

Originally Posted by Charles W. Hull

(unless they said this and I missed it).

They specifically and categorically denied row skipping:

Digital Cinema Society - New Streaming

But I believe that about as much as I believe any of their marketing (i.e. not at all). There are only two possibilities:

1. Canon lied, but they are competent. They are row skipping, because reading all the rows is a huge technical hurdle that no one has yet leaped.

2. Canon is telling the truth, but they are incompetent. This explanation requires you to believe that they are not row skipping because they are years ahead of everyone else in technology. You also have to believe that they botched the implementation so badly that the aliasing is just as terrible as it would be if they were skipping rows.

Personally, I think #1 is more likely.

Originally Posted by Charles W. Hull

Even if they don't it would be possible to have good aliasing performance, given enough processing.

Even if they read the entire sensor and used a cray super computer to downsample, there is still one part that is suboptimal: the optical low pass filter on the sensor is too weak for motion. Of course, it will be worlds better than the 5D2 now, and probably good enough for most applications, but still not as good as a correctly engineered motion camera. For xample, the 20D had a sufficient MTF, but for newer cameras Canon has given in to market pressure for more "sharpness" at the expense of aliasing artifacts.

Originally Posted by Charles W. Hull

I was hopeful that the 7D would perform better with its two processors, that some of the added processing would be used to beef up anti-aliasing, but it doesn't look like it. 7D video looks pretty much like 5DII video.

Agreed.

Originally Posted by Charles W. Hull

I expect it is a thing where if enough of us complain Canon will reduce aliasing in future cameras.

Don't be so sure! All I hear are people loving and praising the aliased images. Dinosaurs like me who actually like natural, realistic looking images instead of the artificial aliased ones are probably just a teeny tiny minority.

In any case, first they would have to conquer the technical challenge of providing alias-free 24p and 60p in a DSLR before they decide if they want to or not. :)

Originally Posted by Charles W. Hull

In the mean time I notice my audience isn't complaining about the aliasing. (But it bugs me!

Agreed.

Originally Posted by Charles W. Hull

But I question whether the explanation is as simple as skipping rows.

I agree that it's not that simple. First there is the aliasing passed by the weak OLPF. Then there is the aliasing from on-chip binning (if there truly is any as Canon says). Then the exacerbation of aliasing by the demosaic algorithm Canon uses. Then more aliasing by the resample up to the video container size (1920x1080).

[Jon Fairhurst]

Daniel, you wrote that "There is no such thing as an anti-aliasing algorithm."

Certainly, if you read all the data from the sensor there is. It's just a 2-d low pass filter. To do a nice job, it should have at least a 5x5 swatch, and larger would be better.

It would be more precise to say "there is no truly effective anti-aliasing algorithm that can be applied to the 1920x1080 output."

I find that there are two types of aliasing effects: on edges, and on patterns. If you get stair steps on edges, filtering can help reduce the aliasing, but the image gets soft. If the problem is a tile roof behind the talent, some area selective filtering could do the trick.

On the other hand, if you get low frequency blotches on a knit shirt, you're hosed. The "algorithm" could be frame-by-frame cloning and painting, if the shot was critical and you had the budget. Ouch!

As to the optical low pass filter being too weak, if we had really good digital filtering off the sensor, one could use a very weak diffusion filter, like a Soft/fx 1/2, and potentially knock down the extreme HF aliasing without smearing the 1080p image.

Personally, my plan is to stick with the 5D2 for filming and to just accept the aliasing, skew and limited color depth. It's more than good enough for Vimeo stuff. The day I have that "special project" that is efficiently scheduled, I'll rent a Scarlet S35, FF35 or similar, once available. While I might want the perfect camera for my personal projects, I don't necessarily need it. I'm just happy to have something to develop my style and skills with.

[Daniel Browning]

Originally Posted by Jon Fairhurst

It would be more precise to say "there is no truly effective anti-aliasing algorithm that can be applied to the 1920x1080 output."

Yes, thanks for the correction. That is indeed what I meant.

Originally Posted by Jon Fairhurst

As to the optical low pass filter being too weak, if we had really good digital filtering off the sensor, one could use a very weak diffusion filter, like a Soft/fx 1/2, and potentially knock down the extreme HF aliasing without smearing the 1080p image.

Agreed.

[Jon Fairhurst]

Now all we need is a DvSLR with proper digital anti-aliasing off the sensor. :)

I figure that such a camera will be available in the standard lineup in 3-5 years.

I also guess that the next step is DvSLR guts in a video body - including aliasing. We should see prototypes at NAB in April. Before long, we will see big sensors with fewer pixels and video appropriate filters in pro video bodies - but these will be expensive, due to the low volumes.

The thing I'm not as sure about is the lens mount strategy. Will we get EF+ lenses that support electronic (rocker) zoom? Or will we get a new video-specific lens mount. I'm hoping for the former.

So, my prediction is:
2010: DvSLR guts in a prosumer video camera body. (And RED cameras.)
2011: Big sensors without aliasing in prosumer video cameras. Maybe more rolling shutter than RED and maybe no RAW. Possibly cheaper than RED.
2012 - 2014: No more aliasing and minimal rolling shutter in the better DvSLRs. RAW video.

I have no inside information. This is my personal speculation. We will see!!!

[Mark Holmes]

How can you make these predictions all the way into 2014? Haven't you seen 2012?

[Jon Fairhurst]

:) :) :)

I forgot to mention that Canon's key engineers will take a rocket ship to another planet where they will safely develop the 2013 and 2014 models. However, sales volumes may be limited...

[Daniel Browning]

Originally Posted by Jon Fairhurst

So, my prediction is:
2010: DvSLR guts in a prosumer video camera body. (And RED cameras.)
2011: Big sensors without aliasing in prosumer video cameras. Maybe more rolling shutter than RED and maybe no RAW. Possibly cheaper than RED.
2012 - 2014: No more aliasing and minimal rolling shutter in the better DvSLRs. RAW video.

Good guess. My hunch is that they will skip the first step (DvSLR guts inside prosumer video camera body) entirely and go straight to big sensors with no aliasing. It's one thing to let a still camera out the door with bad aliasing, that's excusable because it's the stills division. But this will be from the video division, and I don't think they could live with the embarrassment of an aliasing video camera, even though a lot of the market wont mind.

[Jon Fairhurst]

It's that "a lot of the market won't mind" thing that makes me think we'll see DvSLR guts in video cameras. In fact, a lot of the market will gladly fork out the cash.

The company that does this needs to be careful with branding. They need to not confuse it with their broadcast line. And they need to leave some room to differentiate the non-aliasing cameras that will come down the road.

But clearly, many video people will sacrifice aliasing and rolling shutter for a big sensor cam with low noise at a good price. Add XLR audio, zebras, and a video form factor, and it should do well in the market.

And guess what? It will be a video camera that can finally take good stills. ;)

Between RED and the DvSLR revolution, NAB will be VERY interesting this year!

[Kalulu Ngilo]

so is it safe to say that no firmware update can ever fix the alias/moire effects?

[Daniel Browning]

Originally Posted by Kalulu Ngilo

so is it safe to say that no firmware update can ever fix the alias/moire effects?

Yes, that's safe to say. If the hardware was capable of better readout, but just never utilized, then a firmware could fix it. But I think the hardware is already doing the best it can.

[Jon Fairhurst]

I agree that the firmware can't improve the aliasing while scanning the whole sensor. Not only that, but the current Magic Lantern firmware just works in the general purpose CPU. It doesn't work at the hardware scanning level, or the real-time DSP level. Maybe someday it can work at a lower level, but it would be a long leap without proprietary info. (For legal reasons, I doubt that Tramm would touch any proprietary documents with a ten foot pole.)

In theory, one could scan the middle of the sensor without skipping lines and not have any aliasing. Skew could be greatly reduced. Of course, we wouldn't get all that cool DOF, and our wide lenses would become moderate telephotos. That said, it would be a big job for Canon engineers, let alone for reverse engineers.

It would be an awesome option though. I could throw a switch and turn my 200L into the equivalent of a 600mm f/2.8 without aliasing or skew. Whee! I'd become a nature shooter!

I don't expect Canon to provide that option. Those fast, long lenses have good margins. :)