5D Mark III - Plenty Sharp and Still Anti-aliased (video examples)

jcs

Veteran
Watch full screen 1080p to prevent aliasing:

If you are into brick walls:

The method:

1. Shoot Faithful or Neutral (tested so far, custom styles will also work): make sure Sharpening is 0 (this will prevent halos caused by in-camera Unsharp Masking).
2. Shoot IPB (I-frame not quite ready for primetime: either an encoder issue or PPro CS5.5 can't deblock it properly (or some other issue)).
3. Sharpen with a convolution filter. In PPro this is the Video Effect "Sharpen". I have found 25-50 to work well depending on the lens and content. The trick is to increase sharpness but not make noise or other artifacts apparent (including aliasing).
4. Unsharp Masking (USM) can also be helpful (haven't tried used it in PPro CS5.5 as it's not real-time). USM can also be used for local contrast enhancement: http://www.cambridgeincolour.com/tutorials/local-contrast-enhancement.htm. Hopefully CS6 will have real-time USM (it's easy to implement on a GPU).

Some folks believe sharpening in post is cheating, and that the camera should output sharp content directly. The challenge is that as sharpness is increased, aliasing is also increased. To understand better why this is so, take a look at sampling theory: http://en.wikipedia.org/wiki/Nyquist–Shannon_sampling_theorem

For whatever engineering/marketing reason the 5D3 shipped with firmware outputting a highly anti-aliased image (almost no aliasing, even for very difficult content (see videos above)), however some people are used to sharp, aliased content. Keep in mind film has no aliasing, and its apparent resolution is helped with film grain.
When information is encoded in pixels through an anti-aliasing process, it can be sharpened using a convolution filter to increase perceived sharpness at the expense of increased aliasing. For most people, a small amount of aliasing is acceptable (GH2, FS100, F3, etc.). Even the mighty top-end cameras such as the C300 and Alexa have aliasing and moire issues (probably F65 too): http://www.xdcam-user.com/2012/02/c300-moire/. When aliasing and especially moire is recorded, it is hard or impossible to remove in post. Thus the highly anti-aliased footage coming from the 5D3 is a benefit to those shooting clothing and around brick walls, and car grills/detail as with the Bugatti Veyron in the first video above.

So why not just throw a convolution sharpen filter on 5D2 or 7D footage and save the upgrade cost? Because these cameras have baked in aliasing, and the actual frequency information has not been spread into neighboring pixels and can never be recovered (well, not exactly true. Super Resolution uses aliasing information to recover lost frequency information: http://compression.ru/video/super_resolution/super_resolution_en.html, however no commercial tools have incorporated this technology yet). Sharpened 5D2/7D footage ends up with enhanced aliasing artifacts. If moire is present before sharpening, game over. Ultimately sharpened 5D2/7D footage cannot ever look as sharp as sharpened 5D3 footage.

In summary, I am pleased with the video image quality produced by the 5D Mark III. One of my day job tasks is to write CPU and GPU image processing code. I am intimately familiar with image quality issues, and I think Canon has done an excellent job for this initial release. I think it would be helpful for a future firmware upgrade to provide better control for in-camera sharpening: a mixture of USM and convolution. This could save post time for some workflows. I would also prefer the 50Mbps 422 MXF codec to the I-frame 420 codec (VBR: measured at 28Mbps for some of my footage).

Andrew at EOSHD was kind enough to give post sharpening a try, despite his skepticism:

http://www.eoshd.com/content/7608/cinestyle-on-the-5d-mark-iii-and-fixing-softness-in-post#d2
gh2-vs-5d3-sharpness.jpg

He was able to make 5D3 footage comparable to the GH2 (to my eyes, the 5D3 image is a bit sharper, however the cameras are focused slightly differently). He's not really sold on post sharpening, and that's fair, however the 5D3 does indeed encode and store the necessary frequency information for later high-quality sharpening if so desired. The GH2, D800 and similar cameras do not currently have an option to provide little or no aliasing+moire.

Sebastian at Cinema5D also got good results, see his video comparing the 5D3, 7D, and Scarlet-X: http://cinema5d.com/viewtopic.php?f=88&t=39809
From his video you can see the 5D3 closely matches the Scarlet-X for his 2K test (granted the Scarlet-X would do better shooting 4K and performing a filtered downsample). You can also see the 7D cannot be sharpened to match the sharpened 5D3.
 
it is what it is. it looks great without moire and aliasing already.
sharpening cant add fine details.
just accept it. no need to affirm yourself
 
Hi dcloud, you're right, straight out of camera looks excellent, very filmic. However many folks who are used to extremely sharp content feel the 5D3 isn't sharp enough. For example, many feel the GH2, FS100, etc., are sharper and are unhappy that their 5D3 (in their hands or on order) isn't sharper. The above post is to help them realize that the 5D3 imagery contains a great deal more spectral information than is apparent straight out of the camera. It's possible to narrow the frequency content using a convolution sharpen filter to get the look they want. Convolution sharpening doesn't add fine details: it pulls up high frequency information that is already present. Unsharp Masking removes low frequency information.

A DCT or FFT can do something similar in the frequency domain. Consider the DCT-II (http://en.wikipedia.org/wiki/Discrete_cosine_transform), after converting to the frequency domain, we can change spectral coefficients before converting back to the spatial domain. So when you talk about fine detail, you are talking about high-frequency information. If such information is present in the signal, but also accompanied with lower frequency information, we can improve perceived sharpness by effectively increasing the intensity difference between pixels, relative to interpixel derivatives, which is what a convolution filter and USM perform (the results are different, but the concept is the same). I have implemented DCT-II code and have used it to remove high-frequency information for work projects; it could also be used to remove low frequency information, thus performing a sharpen function. When the necessary high-frequency information is not present in the image (signal), it cannot be enhanced with relatively simple filtering. This is why 5D2 and 7D footage cannot be sharpened to get the same results as the 5D3: the high frequency information of the source signal has been lost. Interestingly, aliased signals can provide sufficient information to reconstruct a higher frequency signal, see: http://en.wikipedia.org/wiki/Super-resolution. Fractals can also be used to increase resolution: http://www.ononesoftware.com/products/suite/perfect-resize/?ind.

I hope this information is useful in helping folks get the most out of their cameras.
 
Kinda like this:
AVCHD capture.
surf as is.jpg
Straight out of camera, surf blown out.

surf recovered.jpg
Highlights recovered. No sharpening tool applied. Took 10 secs.
 
Adding contrast gives the perception of sharpening too, so shooting with cinestyle and crushing the blacks could work quite well.....damn wish I had a camera to play with.
 
Right- sharpening is effectively increasing high-frequency contrast. Pulling down blown highlights can do the same (for any color components that weren't clipped). Local Contrast Enhancement is also a helpful tool: http://www.cambridgeincolour.com/tutorials/local-contrast-enhancement.htm. Give it a spin with your NLE's Unsharp Mask if you haven't tried it before (works more on mid and lower frequencies). HDR-style tone-mapping can also help improve perceived detail (local adaption: can be use with any image, not just HDR): http://en.wikipedia.org/wiki/Tone_mapping, though care must be taken to prevent an artificial look (unless that is the goal). This is something to try in Photoshop CS6 (free beta is out): has much better video editing support. PPro's "Shadow/Highlight" effect looks similar (I tend not to use any non-GPU effect/filter though. Hopefully CS6 will GPU accelerate all the filters).

Looks like you settled the debate in the other thread. Thanks for the film grain sequence.
 
c300 probaly use APS-C crop of canon 5dmark iii sensor ...
when u upscale c300 to full frame u got 22.3 MP...(2.14158238 * 8MP * 1.33 = 22.7864365)
all companies do that ...
design one sensor technology and the upscale or downscale size and deactivate some options for cheaper or photo cameras ...
image from 5d iii sensor is probably perfect 4 MP and then canon use some black magic to do worst possible downscaling to 1080p ...
c300 has huge profit margin ...
if canon does no line skipping and read whole sensor with pixel binning only way to get such low resolution HD is on purpose to protect video division...
 
Looks like you settled the debate in the other thread. Thanks for the film grain sequence.

hehe, nope just stoked the flames. It will keep burning until the next new thing comes along for everyone to fight over.

You know much about lineskipping John? The 22.3MP sensor is an exact 3x cropped downscale to 1080p. I took a full frame MKIII raw still, downscaled and cropped to a 1080p high quality jpeg and it was around 500k. 24 x 500k is about 90Mbit/s same as the ALL-I codec so does that mean it's not lineskipping like the MKII?
 
Last edited:
The Bayer array makes it more complicated. I guessed they were binning green in a 2x2 pattern reading 4 or 5 pixels yielding 1920x1080pixels, and red and blue was binned in a 3x3 pattern yielding 960x540. This reduces the resolution of Y somewhat, however it would be fast for the hardware. Now that I have the camera, it's still not clear what they are doing, but they are definitely not line skipping, otherwise there is no way they could be doing such a good job with anti-aliasing. In the unlikely event they are reading the entire sensor and downsampling with the Digic 5+, future firmware updates could unlock much higher quality.
 
Given that digic 5 is apparently 17x faster than digic 4 do you think it's possible that it could downsample 24fps on the fly?
 
I think you meant downsample to 1080 and 720p? I think the limiting factor may be sensor readout speed. If not, then we'd need to know 17x faster that what baseline performance. Resampling can be very fast in GPUs, fully hardware accelerated. Don't know what they built into the Digic 5+.
 
...Some folks believe sharpening in post is cheating, and that the camera should output sharp content directly...

jcs - thank you so much for clarifying the technical details surrounding the balance between moire and aliasing. A really well-informed and well-thought out argument. A couple of thoughts -

I don't think sharpening in post is cheating, but I do think it takes extra time - and time is money.

I guess, for me, if the choice is (leaving color issues aside):

1. Old Canon - unacceptable moire, sharp enough for most people
2. New Canon - minimal moire, post shoot sharpening required
3. Brand X - acceptable moire, acceptable aliasing, no post sharpening required

I'll take option 3 (unless the full frame aesthetic is an absolute must have, then I'll rent).

No offense to anyone intended, but I just don't have the time for a camera that forces me to take an extra step in post, even it takes 10 seconds (or forces me to either buy a monitor or lay down on the beach like poor Mr. Bloom :-))

photo-2-6.jpg

Cheers,

Bill
 
Definitely an improvement over the 5Dmark 2, but still looks pretty soft in those wides for anything that is infinity focus. Aliasing and moire is improved, but not gone. I really have been spoiled with my hacked gh2 lol.
 
brunerww- if you have a camera (or cameras) that work better for you than the 5D3- makes perfect sense to use them. Dragging a sharpen preset to a sequence with all clips selected takes about 3 seconds, so if 10 seconds is your threshold, you are still good to go with this workflow :smile:

PaPa- the wide soft comes from the Canon 16-35mm F2.8L II: it's not a sharp lens. I could probably get it sharper by stopping it down, etc. Nikon and Zeiss ($$$) have sharper ultrawides, which would work better. Where did you see moire in those videos other than perhaps one shot of the Bugatti Veyron grill. Out of the camera before sharpening, there was no aliasing or moire. I did sharpen in many clips to the point of a little aliasing. We can't have it both ways: razor sharp and alias/moire free, until perhaps we get to 4k and beyond (and high resolution display systems). As noted, adding a pleasing film grain can help with perceived resolution in the event the content creator requires zero aliasing or moire. A more technical solution for the perfectionist could be to apply a bilateral filter, which removes high frequency content but preserves edges, tuned to remove frequencies in edges near the alias limit (Nyquist), followed by a sharpen (or edge sharpen).

Emanuel- no problem. These concepts can be applied to any camera or imagery.
 
brunerww- if you have a camera (or cameras) that work better for you than the 5D3- makes perfect sense to use them. Dragging a sharpen preset to a sequence with all clips selected takes about 3 seconds, so if 10 seconds is your threshold, you are still good to go with this workflow :smile:

PaPa- the wide soft comes from the Canon 16-35mm F2.8L II: it's not a sharp lens. I could probably get it sharper by stopping it down, etc. Nikon and Zeiss ($$$) have sharper ultrawides, which would work better. Where did you see moire in those videos other than perhaps one shot of the Bugatti Veyron grill. Out of the camera before sharpening, there was no aliasing or moire. I did sharpen in many clips to the point of a little aliasing. We can't have it both ways: razor sharp and alias/moire free, until perhaps we get to 4k and beyond (and high resolution display systems). As noted, adding a pleasing film grain can help with perceived resolution in the event the content creator requires zero aliasing or moire. A more technical solution for the perfectionist could be to apply a bilateral filter, which removes high frequency content but preserves edges, tuned to remove frequencies in edges near the alias limit (Nyquist), followed by a sharpen (or edge sharpen).

Emanuel- no problem. These concepts can be applied to any camera or imagery.

I saw it in the shot of the jacket. It was a mix of internet compression and moire.
 
brunerww- if you have a camera (or cameras) that work better for you than the 5D3- makes perfect sense to use them. Dragging a sharpen preset to a sequence with all clips selected takes about 3 seconds, so if 10 seconds is your threshold, you are still good to go with this workflow :smile:

PaPa- the wide soft comes from the Canon 16-35mm F2.8L II: it's not a sharp lens. I could probably get it sharper by stopping it down, etc. Nikon and Zeiss ($$$) have sharper ultrawides, which would work better. Where did you see moire in those videos other than perhaps one shot of the Bugatti Veyron grill. Out of the camera before sharpening, there was no aliasing or moire. I did sharpen in many clips to the point of a little aliasing. We can't have it both ways: razor sharp and alias/moire free, until perhaps we get to 4k and beyond (and high resolution display systems). As noted, adding a pleasing film grain can help with perceived resolution in the event the content creator requires zero aliasing or moire. A more technical solution for the perfectionist could be to apply a bilateral filter, which removes high frequency content but preserves edges, tuned to remove frequencies in edges near the alias limit (Nyquist), followed by a sharpen (or edge sharpen).

Emanuel- no problem. These concepts can be applied to any camera or imagery.

How is sharpening a 5D III image better than sharpening a 5d ii with AA filter image if both only resolve under 800 lines in the raw file? Is the argument that the 5d III resolves 900+ lines, but appears soft due to no aliasing? Because if that were the case, the raw clips would still have more visible detail... and that isn't apparent in the clips being posted. Sharpening exaggerates detail that is already there. I'm not seeing that initial detail.
 
Last edited:
Back
Top