Aliasing
 

While this has been discussed in different threads aliasing is still the number one issue for me. I don’t own the 5DMK2 so my conclusions relay on posted material. First I thought it is only a problem of wrong downscaling for the web (vimeo, flash, QT). But then I started to look for aliasing in raw footage and it is already there. Also it can be found in various footage, so I don’t think everyone is doing something wrong while downscaling. I’m engineer my self and understand the problem Canon was facing with the 5DMkII. They had to design the OLPF (optical low pass filter aka anti aliasing filter) for the 21 Mega pixel resolution of the Sensor rather than for the 2MP used for HDTV. For properly sampled video they had 4 options (IMO):

1. Read out the whole 21 mega pixel 30 times per second (630MP/S) and downscale it to 1080p30 using e.g. bilinear filtering. Even if the CMOS sensor could be read out that fast (I highly doubt that) the resulting data stream would most probably be more than the 5DMKII chipset can handle.

2. Read out only the center of the CMOS sensor (aka region of interest or windowing). While this is very easy to implement it has a lot of drawbacks. First of all it introduced a crop factor like on APSC Cameras. But in this case it will multiply the focal length of your lens not only by 1.5 as on APSC sensors but by 10 (yes then). While this is great news for wild life shooters, it is absolutely not usable for anything else but filming a close up of the moon with your 70mm lens.
Compared to the full readout the sensitivity will drop by at least 3 stops.
Also you will loose much of the shallow DOF. It would be slightly better than a 2/3” sensor.

3. There are CMOS sensors which can combine adjacent pixels (aka binning) and read it as one single large pixel. In the case of the 21MP sensor of the 5dmk2 a 3x3 pixel binning would reduce the amount of data by a factor of 9 to a moderate 70 MP/S. This data rate could be easily handled by the cameras chipset. Unfortunately the 5dMk2 sensor seems not to support binning.

4. Most CMOS sensors can skip rows and pixels to reduce the resolution. This is normally used for preview functions where not the whole resolution is needed. It seams the Canon engineers use this method to reduce the amount of data for the Video mode. The problem is; when skipping e.g. every second pixel the OLPF should be twice as strong (it now should blur an infinite small point in the image to a disc of the diameter equal to the distance of 3 sensor pixels). While such an OLPF would produce clean video the still resolution of the sensor would drop from 21MP to 5MP (not a good thing for a professional DSLR). So Canon decided to go for sharp 21MP stills over properly anti aliased video.

Until either the chipset and imaging sensors are fast enough to handle 30 full sensor readouts per second or the sensor supports pixel binning there will always be a tradeoff between still resolution and proper anti aliased video.

Once captured it is not possible to completely remove the aliasing in post. Or at least I do not see how this should work. The only solution I can imagine is to add some sort of optical diffusion filter.

This is good news Josh. Can it be that simple? This should be investigated more deeply.

Solution 2. is what RED is doing when shooting 2k, quite annoying as you point out.

If you're right (Canon using Solution 4.) that's disappointing (but that's what I guess from user reports too :/) since we could have 10 times better image quality (and noise-free video @6400 ISO) if bilinear downsampling was used...

Do you have some link to raw footage showing aliasing artifacts ?

Have a look to Video Samples 2 from Digital Photography Review
The shop door in the background shows some aliasing. Yes I know its shot with a preproduction model but I don’t think there has been any significant improvement to the production model.

Dan Chungs Uganda video on Vimeo shows aliasing at 2:30 on the woman’s shirt. Dan can you confirm this aliasing is also in the raw data?
P.S. if I was told to give an interview in front of a 5DMk2 I’d wear a nice black and white checked suit :-)

Red doesn't alternate lines, so it's not as much of a problem for them. We don't know if Canon is skipping lines to reduce data or the read/reset isn't fast enough.

I'm not sure if they are just row skipping (1/3rd the pixels) or row+col skipping (1/9th?). The effective "fill factor" must be very low, perhaps less than 10%.

It's hard to decide what I dislike the most about my new 5d2. Aliasing is up there very high, but I have to balance that against all the other failings:

• Aliasing, moire, stair-stepping, jaggies, etc.
• Lacks even the most rudimentary manual controls
• Sets shutter speed based on the 1/focal length rule
• 30p
• Compression artifacts
• Heavy and ugly noise reduction even in "NR: off" mode, poor demosiac.
• Stops at 4GB instead of writing a new file.
• No audio meter, control, or headphone monitor.
• No live HDMI

It seems as if Canon got just about everything wrong; the 5d2 does very poorly in just about every possible metric for measuring the quality and utility of a video camera, except one: sensor size.

Unfortunately for me, this means I will have to continue lugging around my XH-A1 in addition to the 5d2 until Canon releases a free firmware upgrade (it is to laugh).

I just got Tiffen Soft #1 and Diffusion #3 filters. On a 50mm f/1.8 Nikon, the Diffusion filter definitely removes all aliasing. The Soft filter seems too weak. If I had to guess right now, a Soft #2 would be the target.

I need to do more testing before I'm certain about that though...

Thanks, Jon; I appreciate your testing. I will pick up a Soft #2.

That sounds very interesting, can you put some footage up?
But Tiffen also offers Diffusion #2 and #1 - why would you assume soft #2 would be strong enough?

I wonder... there's already a filter to deal with the d90's aliasing in post - from the plugin developer (Too Much Too Soon Free Plugins for Final Cut Pro and Final Cut Express):

"The D90 rescaler plugin works because the D90 really captures an 800p image and scales it to 720p, and the way it's done makes it possible to scale it back up almost losslessly to 800p and then back down using a better algorithm."

maybe something similar can be worked out for the 5D? If we can figure out how many sensor lines are being skipped it should give us a target resolution to try the same process.

EDIT: hmm, maybe not - taking a 16x9 crop of the full resolution (5616 x 3744 pixels) gives us 5616x3159. Sampling every 3rd horizontal line would give us 1053 so maybe they're not doing any additional downsampling which could be re-done in post as with the d90.

You're right that it's an assumption right now. I haven't had a chance to test with it.

From what I've seen the Soft filters give an overall blur, while the Diffusion filters bounce the highlights around for more of a highlight blur effect. How they do this, I have no idea. Somehow they keep clean details in the blacks, but smear only the brightest aspects of the picture. That makes me surmise that the Soft filter is the better choice for anti-aliasing. But I don't, in fact, know that a #2 is strong enough. But from my quick tests, we can scratch the #1 off the list.

Thanks for that interesting information!
The downside is, that these filters simply are not available in Europe - so I´d have to buy them in USA. I would really have to be sure to buy the right one - so please keep us informed if you find any information or even RAW-Video using one of these filters.

I will try to get my hands on one of these german filters:
-Zeiss Softar
-Heliopan DUTO
-B&W Soft Pro
Unfortunately they are just available in 2 versions.
As B&W is available in USA too, if anyone got experience with them, please post! :-)

Unfortunately, I won't be able to do any tests for a few weeks. My son took the camera with him as he returns to college, where he will shoot the pilot of ~ The Murder of Dirk Snowglobe - A Not Dead Detective Series ~.