GH1 Scaling - Fiction..... ?

D

Daniel L.

Well-known member
I'm trying to find out where this idea that the GH1 is scaling video comes from, and if it's true. I keep bumping into people insisting this is a fact but it seems extremely unlikley to me.

I would like to be wrong, if it's really scaling then that's fantastic. Where does that come from? If anybody has a source with some objective analysis I would appreciate a link.

Aliasing and moire are both present on the GH1 videos. These are exactly the same artifacts as you see on any camera skipping rows (i.e. Mark II).

Maybe it's not the same.. Maybe it's just really poor filtering, or a very sloppy algorithm... But there is no reason to believe this. The most practical approach to implementing a video function for these video hardware deficient cameras is to subsample, or skip photosites.

The reason for my confusion is that I observe this problem all over the place. It's in nearly every video. Here are some crops from a raw GH1 clip: <edit> Err.. this example is not from a raw clip, details two posts below </edit>

1.jpg

2.jpg



More than likely this discussion will follow the trend and turn into this camera vs that one and a mob of people yelling at each other completely off topic. However, if anybody has any information concerning scaling please let me know.

A ISO 12233 chart would be very helpful, but I have not been able to locate one yet. If someone would shoot one and post it here it would answer a lot of questions.
 
Last edited:
The only time I've seen what you are showing is when I've accidentally rendered converted video to the wrong resolution and the problem is caused by the rendering tool not doing a quality resample. When that happens or when the video is recompressed too much, I see what you posted. Did you get these snaps from recompressed video or the original MTS files?

Mike
 
I just found the video I referenced, I'm mistaking. It was not raw, but a Quicktime with Cineform conversion for pulldown removal. This invalidates the crop examples above... I'm looking over some MTS files to see if I can find the conditions that would demonstrate if this is a problem or not.

I have seen a pattern of the same aliasing on numerous videos, that's how it got my attention in the first place. You are right that this can be introduced with a poor post workflow.

If someone is interested in shooting a chart it can be found here: http://www.graphics.cornell.edu/~westin/misc/res-chart.html
 
Last edited:
Take a look at the GH1 video here:

http://www.vimeo.com/5623066

The 5D Mark II version is linked below the video. Watch both and you'll see that there is significantly less moire in the buildings just to the right of center on the GH1: 5D2 has more moire/aliasing. That might suggest that the GH1 is scaling BUT the fact that the GH1 still has some/any moire on an image that should be downsampled by a factor of 2 makes me wonder. I'd think if the image was scaled, there would likely be no aliasing at all! So I really don't know what to think. It's a known fact that you must apply some antialiasing to get downsampled images without moire, so maybe it is scaled but for speed, no antialiasing is done in the downsample. That might explain what little aliasing is there.

Mike
 
I really hate to bring the 5D anywhere near this topic. This is not about which camera is more prone to problems... it's about finding the truth. Don't get me wrong, I'm not suggesting that's your intent... Just a disclaimer clarifying my own position.

Say hypothetically both cameras are subsampling photosites: Since the 5D sensor is twice larger in area than micro four thirds, the spacing in between sampled photosites would be increased. This would result in more aliasing & moire, precisely how we observe in the Mark II sample.

My point is that just because the GH1 shows less does not mean it's not subsampling photosites. This is no justification for a scaling claim.

I don't think we will get very far comparing transcoded video samples anyway...
 
Daniel L. said:
I really hate to bring the 5D anywhere near this topic. This is not about which camera is more prone to problems... it's about finding the truth. Don't get me wrong, I'm not suggesting that's your intent... Just a disclaimer clarifying my own position.

Say hypothetically both cameras are subsampling photosites: Since the 5D sensor is twice larger in area than micro four thirds, the spacing in between sampled photosites would be increased. This would result in more aliasing & moire, precisely how we observe in the Mark II sample.

My point is that just because the GH1 shows less does not mean it's not subsampling photosites. This is no justification for a scaling claim.

I don't think we will get very far comparing transcoded video samples anyway...
Click to expand...

I agree, it isn't about 5D versus GH1. I brought it up because it is a comparison and we know the 5D2 skips lines and doesn't scale and because you mentioned that in the OP. I think your logic is faulty though. Both videos have the same angle of view and since both videos cover that same angle of view with the same number of pixels (1920 x 1080), it makes no difference how large the sensor is. Relative to the scene, the pixel spacing is the same.

Mike
 
I think it's important to understand where the moire & aliasing come from in the first place. They are two different beasts that result from similar conditions. The source of the moire is high spacial frequencies above Nyquist that are not properly sampled by the sensor. This is due to the gaps introduced by subsampling photosites. Cameras have optical lowpass filters to eliminate high spacial frequencies, however in this case they are out of sync. Using a filter designed for HD on either camera would severely hinder the resolution of the stills capability. So that was never an option for either of the companies... and we are left to deal with it on our own. (diffusion, etc..)

Since both cameras are sampling the same number of photosites, but with different area there will be differences in gaps between photosites. The greater the distance between photosites, the lower spacial frequency is necessary to cause these problems.

In the end this has nothing to do with field of view, it's just helpful when comparing two sources side by side. You can find high spacial frequencies regardless.

There is more to it than this, but keeping this short and simple as possible.
 
I have a simple theory that the 5D doesn't do scaling because it has a 21 megapixel sensor. The grunt required to do the scaling would be too high for current onboard image processors.

I also don't believe the GH1 scales either. It might just have better image processing to remove the moire & aliasing.

A 21 megapixel sensor resolves a heck of a lot of detail and the GH1's sensor is known for having incredible resolution for it's class, something which the DPReview.com raw tests picked up on. So I guess if they're scaling the images, both camera's video still frames would look better than they currently do.

Have you tried taking a still photo and a video of the same scene and scaling the still photo down to the video frame size (1920x1080)?

I tried this and not withstanding the obvious compression artefacts of AVCHD, there is a world of difference. There is just not as much detail left in the video frame compared to the scaled still frame and that is why I don't think the GH1 scales.

Once these sensors & processors are capable of outputting 14 megapixels at 24fps, and scaling it down to HD we will have a serious jump in image quality. But it's a long way off (in my opinion).

MJPEG frame capture, 1/50 F2.0 720p
movie720.JPG


12 megapixel still in 16:9 mode, 1/50 F2.0 scaled to 720p
still720.JPG
 
Last edited:
Daniel L. said:
Since both cameras are sampling the same number of photosites, but with different area there will be differences in gaps between photosites. The greater the distance between photosites, the lower spacial frequency is necessary to cause these problems.
Click to expand...

True, but once you've gone beyond the radius of the low pass (AA) filter, it really won't matter as the filter will not be doing what it was designed to do anyway. Are you saying that since the GH1 will have to skip less lines that the AA filter may still have some weak effect on neighboring photosites in the skipped array? That's possible but it may be impossible to distinguish between that type of hardware aliasing and aliasing caused by downsampling.

What I'm saying is that if you simply downsample an image without any antialiasing logic, you're going to end up with the same type of moire that you'd get if you remove the AA filter. High quality downsampling involves an antialiasing step that is separate from the downsampling of the pixels. Unfortunately the antialiasing is many times slower than the downsampling itself so if they (either the 5D2 or GH1) are doing a quick downsample that doesn't include the anti-aliasing step, it'll alias just as bad as skipping pixels because the end result of a weak (or no) antialiased downsample amounts to basically the same thing as skipping pixels. I guess in that same vein, it depends on what you define as "scaling".

Bottom line: just because we see moire, we can't rule out either method.

Mike
 
Last edited:
I heard somewhere that the GH1 has a weaker AA filter than other cameras which would explain the extremely sharp RAW output (with a good lens).
 
commanderspike said:
I have a simple theory that the 5D doesn't do scaling because it has a 21 megapixel sensor. The grunt required to do the scaling would be too high for current onboard image processors.

I also don't believe the GH1 scales either. It might just have better image processing to remove the moire & aliasing.

A 21 megapixel sensor resolves a heck of a lot of detail and the GH1's sensor is known for having incredible resolution for it's class, something which the DPReview.com raw tests picked up on. So I guess if they're scaling the images, both camera's video still frames would look better than they currently do.

Have you tried taking a still photo and a video of the same scene and scaling the still photo down to the video frame size (1920x1080)?

I tried this and not withstanding the obvious compression artefacts of AVCHD, there is a world of difference. There is just not as much detail left in the video frame compared to the scaled still frame and that is why I don't think the GH1 scales.

Once these sensors & processors are capable of outputting 14 megapixels at 24fps, and scaling it down to HD we will have a serious jump in image quality. But it's a long way off (in my opinion).
Click to expand...



Any reason why the Mjpeg is brighter? I balanced them in PS, and that Mjpeg is performing quite well and they get pretty close to one another. Now a 1080/24 of that same shot scaled scaled to 720 would look even better!

.
 
At first I thought I'd shot at different settings but both were shot at:

ISO 320
1/50
F2.0 (Canon FD 50mm prime)

Nope, it just seems that the MJPEG codec produces a brighter image. I dislike it. If you save the photos on the previous page and blow them up in your image editor you'll see that the scaled still shot is much much cleaner.

It has none of the over-sharpening, blown highlights and JPEG compression artefacts of the MJPEG frame capture.

Indeed a scaled 1080p shot to 720p might indeed look better, but a scaled still to 1920x1080 would look better than a still frame from the 1080p video mode.

However, all this is kinda irrelevant in a way. When it's moving it looks a lot better.

And I still have no proof as to if the GH1 scales or not :) but I doubt it does judging from this little comparison. If it does... it does so in a quick and dirty way.

Neither MJPEG or AVCHD at the bit rates the GH1 is outputting are exactly what you'd call 'clean'.

This is one respect where the 5D wins out.
 
commanderspike said:
At first I thought I'd shot at different settings but both were shot at:

ISO 320
1/50
F2.0 (Canon FD 50mm prime)

Nope, it just seems that the MJPEG codec produces a brighter image. I dislike it. If you save the photos on the previous page and blow them up in your image editor you'll see that the scaled still shot is much much cleaner.

It has none of the over-sharpening, blown highlights and JPEG compression artefacts of the MJPEG frame capture.

Indeed a scaled 1080p shot to 720p might indeed look better, but a scaled still to 1920x1080 would look better than a still frame from the 1080p video mode.

However, all this is kinda irrelevant in a way. When it's moving it looks a lot better.

And I still have no proof as to if the GH1 scales or not :)
Click to expand...


Since were still in the 1080/720 realm it makes a world of difference for that frame size to have a camera with a large dense 12mp sensor feeding it. Granted I would love to have a bigger frame size come off that sensor. However I'm happy that we are going in the right direction..


.
 
stav1606 said:
I disagree with that. It does not win because it takes a long time to readout the sensor and thus it has a lot more jello. Nothing comes without a price.
Click to expand...


The problem is; were like kids, and we are getting very spoiled really fast.

The S/N ratio is better on the GH1 then any video camera any of us could afford prior to this; let alone the high frequency resolution in its core senor image.


We're getting a Playboy playmate at a trailer park price tag...

That's a damn good deal!

.
 
I'm not certain I like the analogy!

I'm thinking the GH1 does scale; I'm not getting much aliasing problems. Moire is a totally different thing. It probably doesn't use some super duper sub-sampling process when it scales though.
 
yslee said:
I'm not certain I like the analogy!

I'm thinking the GH1 does scale; I'm not getting much aliasing problems. Moire is a totally different thing. It probably doesn't use some super duper sub-sampling process when it scales though.
Click to expand...

I think you hit the nail on the head yslee. I've noticed that many armchair experts tend to forget that all scaling is not created equally.

Scaling from full sensor resolution (4K) down to 1920x1080 24 times a second is a big task. No one should expect a $1500 camera to be able to pull off the same level of scaling quality as Photoshop, XNview or any other desktop image processor (where speed is far from the primary factor), I don't.
 
Slight correction there, Illya. The camera itself will probably sell for $750 or less. $1500 includes the expensive lens.

For those who have a program that offers different scaling processing, try taking a photo in full 12mp mode, and then scale it down using the following:

1) Nearest Neighbour (as good as dropping alternate pixels)
2) Blinear
3) Bicubic
4) Lanczos
 
Of course! If the GH1 scaled it wouldn't do so to the same sophistication or quality as Photoshop, that goes without saying. Though the point I am making is that I have to see proof that scaling is done at all.

Anyway - none of this matters. The GH1 delivers a good image without some of the issues of the 5D MII, if I wasn't so interested in the technology I wouldn't care how it got to the end result as long as it was good.
 
You must log in or register to reply here.
Back
Top