Solution for pulsing I-frames?

[cbrandin]

The pulsing shows up as a huge P frame that appears immediately before an I frame, and evidentially it contains mostly garbage data. It's generally much bigger than any of the P frames. I frames are frames that are "fresh" in that they don't depend on any prior frames; they are complete images. P frames are "delta" frames in that they only contain the difference between the current frame and the previous one, which is more efficient than having every frame be a complete image. Generally I frames are bigger than P frames (twice as big, or better). Having a huge P frame makes no sense because an I frame is the entire image, which is about as uncompressed as you can get. P frames that are bigger than I frames effectively defeat the purpose of having P frames in the first place. Typically, the ratio of I frames to P frames is the number of frames in 1/2 second of video, so for 60 FPS video that would be one I frame for every 29 P frames.

The main problem with the "pulses" is that they are apparently filled with garbage, not image data, which is not a good thing.

Chris

 

[svecher]

I don't understand what the pulsing is. Could someone post an example, maybe frame without and a frame with it?

EDIT 2: This is probably related to 24pN not working properly and causing camera freezes. These could be those frames that are suddenly recorded with too much information causing the camera to stop recording, error out, and/or freeze. Has this "pulsing" occurred with any non 24pN footage?

I have yet to see them in 60i footage.

 

[cbrandin]

Right, I've only ever seen them with the Native 24p/25p option checked. Sorry if my reference to 60 FPS caused confusion.

Chris

 

[Martti Ekstrand]

I have 'the pulse' in footage shoot last year with a totally unmodified PAL model. It can appear on fine-grained detail areas with low contrast and low light, around 10 - 25% black. Like darker parts of foliage. Never been a biggie for me as I typically have something in the foreground keeping the audience's attention and seldom stays on a take for more than 4 seconds. Once I drew a garbage matte around such a area and applied a average frame filter to smooth away the pulse.

 

[wturber]

The pulsing shows up as a huge P frame that appears immediately before an I frame, and evidentially it contains mostly garbage data. It's generally much bigger than any of the P frames. I frames are frames that are "fresh" in that they don't depend on any prior frames; they are complete images. P frames are "delta" frames in that they only contain the difference between the current frame and the previous one, which is more efficient than having every frame be a complete image. Generally I frames are bigger than P frames (twice as big, or better). Having a huge P frame makes no sense because an I frame is the entire image, which is about as uncompressed as you can get. P frames that are bigger than I frames effectively defeat the purpose of having P frames in the first place. Typically, the ratio of I frames to P frames is the number of frames in 1/2 second of video, so for 60 FPS video that would be one I frame for every 29 P frames.

The main problem with the "pulses" is that they are apparently filled with garbage, not image data, which is not a good thing.

Chris

I don't think the extra large P or I frames (sometimes they are I frames) are the cause of the pulsing. From what I've observed, the pulse is actually a lower quality bump. It is easier to see it in things like low contrast textured walls. What you will see is that at the pulse, some of the wall detail goes away and it then slowly builds over the following frames. I think what is happening actually pretty much the opposite of what you described.

I think what happens is that the I frame is displayed and then each subsequent P frame adds more data. More P frames on a static shot results in a building of detail. Finally, at the end of the GOP, you have the most detailed image yet. Then whammo - you get a new I frame with less detail - and that's the visible pulse.

Higher bitrates and smaller GOPs seem to help, though they don't eliminate the problem. Pulsing seems to be significantly reduced when using a 10 frame GOP for 720p and a 12 frame GOP for 1080p24. Part of the reason for this may be that there is less image building with the smaller GOP. And while I haven't verified this, it does seem that with smaller GOP settings, the I frames tend to be a lot larger than the P frames.

 

[GMC]

I don't think the extra large P or I frames (sometimes they are I frames) are the cause of the pulsing. From what I've observed, the pulse is actually a lower quality bump. It is easier to see it in things like low contrast textured walls. What you will see is that at the pulse, some of the wall detail goes away and it then slowly builds over the following frames. I think what is happening actually pretty much the opposite of what you described.

I think what happens is that the I frame is displayed and then each subsequent P frame adds more data. More P frames on a static shot results in a building of detail. Finally, at the end of the GOP, you have the most detailed image yet. Then whammo - you get a new I frame with less detail - and that's the visible pulse.

Higher bitrates and smaller GOPs seem to help, though they don't eliminate the problem. Pulsing seems to be significantly reduced when using a 10 frame GOP for 720p and a 12 frame GOP for 1080p24. Part of the reason for this may be that there is less image building with the smaller GOP. And while I haven't verified this, it does seem that with smaller GOP settings, the I frames tend to be a lot larger than the P frames.

I agree totally to your first explanation, the I frames don't seem to have enough bandwidth. I am not shure about the smaller GOP part. But I hope that the research on Vitaliys end continues.

 

[alignment1]

EDIT 2: This is probably related to 24pN not working properly and causing camera freezes. These could be those frames that are suddenly recorded with too much information causing the camera to stop recording, error out, and/or freeze. Has this "pulsing" occurred with any non 24pN footage?

I'm using 'jobless' old 'D' settings (now 'C'), w/ 24p checked...I've never had the camera freeze, stop recording error out, etc.. I've been shooting quite a bit this way with no issue, (fingers crossed)rogar-Kriz(DBG):.

As I said in earlier post, I went back to old footage and found it there as well..?!?

What's the deal!?!? How come I never noticed this before?? I even went back to old footy from months ago- (way PRE-GH13) days- factory firmware 720p60 tripod shot of mountain ...PULSE...PULSE...PULSE...

 

[Nitsuj]

When the GH1 first came out I did a lot of testing because I noticed it. I found it shows up much easier when there is a lot of detail, motionless, and a contrast in the image.

Here is the old test I did.

http://exposureroom.com/60839febfaa540f8964174db9c269468/
You can really see it good on HD/full screen

I will repeat it today with the gh13 at the same location to see how much difference there is and post it up later. It has been a long known issue. That is why in high detail shots you want to go soft or avoid due to the long GOP on these HDSLRs.

[edit] Old solutions were
1. Avoid high detail like forests when possible.
2. If have detail avoid high contrast (evenly light if possible)
3. Soften up the image
4. Provide movement in the shot
5. Focus on subject in foreground and blur out the detail with long lens

 

[cbrandin]

The pulse is definitely a bug, and not just the build-up of data. In some cases it simply shows up as two I frames in a row (corresponding to a GOP of 1) with no Pframes between them. As far as I know, this should be impossible.

Chris

 

[cbrandin]

Can someone with a good "pulsed" clip run a little test? Locate the pulsed frame with StreamEyE. Then examine the clip frame-by frame around that location to see when the visual change occurrs. Does it start before, at, or after the pulse? Is the pulse when it gets worse or better? Or, when does the image get worse, and when does it get better?

This way we'll be able to determine whether the pulse is trying to fix something or whether it is just a bug that degrades the image for no reason.

The idea that it puts a bunch of detail in the stream right before the image is to be replaced with an I frame anyway seems rather odd to me.

I would do this test except I don't have any clips that show an obvious visual change during pulses.

Chris

 

[mrduck]

It seems that pulsing in PAL gh-1 was present from the start - only on some recording scenes. It seems to be totally absent from NTSC cameras in 60i and present if 24p mode is selected... My idea is that gh-1 was engineered for NTSC, 60i, and I/P frames were allocated precise amounts of bandwidth. Going PAL they were not able to allocate bandwidth evenly, due to a limited hardware encoder or a wrongly written bandwidth allocator formula or parameters... In case of a limited hardware encoder (buffers, for example) we can do nothing... in case of a wrongly written formalus is difficult but not impossible, for wrong parameters we should find their location in the firmware first.... and test test test...

The pulse is definitely a bug, and not just the build-up of data. In some cases it simply shows up as two I frames in a row (corresponding to a GOP of 1) with no Pframes between them. As far as I know, this should be impossible.

I don't think two consecutive I frames I are forbidden....

 

[cbrandin]

Yes, but the GOP length in the GH1 is a fixed number, implying that it is intended to operate with a fixed GOP.

Chris

 

[cbrandin]

Are you saying that the GH1 can produce pulses even when Native 24p/25p is not being used? Or, when the firmware has not been patched at all?

Chris

 

[mrduck]

Are you saying that the GH1 can produce pulses even when Native 24p/25p is not being used? Or, when the firmware has not been patched at all?

Chris

this has been reported in this thread, read the previous page, from PAL users. I have pulses too, but I only go 720... and I recognized pulsing only when I saw my images on a 50" screen .... I didn't notice before...

 

[cbrandin]

Do the pulses show up in EyeStream, or are we assuming that they are based on image examination?

Chris

 

[wturber]

The pulse is definitely a bug, and not just the build-up of data. In some cases it simply shows up as two I frames in a row (corresponding to a GOP of 1) with no Pframes between them. As far as I know, this should be impossible.

Chris

Again, you are attributing the visual "pulse" with some of the interesting spiked I and P frames. But it seems very likely that the real issue is that the I frame simply isn't providing a high enough initial quality when the displayed frame is generated completely from the single I frame.

When you go through a clip frame by frame, you can see that the first frame is somewhat lower quality and that the image quality improves as the clip steps through each subsequent P frame. When there is movement in the image, the extra data from the P frames is used almost exclusively to describe that motion. But when the image is static, the extra data from the P frames improves on the quality of the static image. You can watch the detail build as the clip steps through the GOP. So when the next GOP begins with a new I frame, we see an instant quality drop - and that's the "pulse" we are seeing. You can see this pulse on clips that don't have a spiked last P frame or unusually large "spiked" I frames.

The spiked I and P frames you are referring to seem to be connected to the camera freezing on static, high detail shots. They don't seem to be the cause of the pulsing.

 

[wturber]

Can someone with a good "pulsed" clip run a little test? Locate the pulsed frame with StreamEyE. Then examine the clip frame-by frame around that location to see when the visual change occurrs. Does it start before, at, or after the pulse? Is the pulse when it gets worse or better? Or, when does the image get worse, and when does it get better?

Chris

The visual "pulse" is the I frame. I checked this out this morning with about a dozen clips. The pulse is always at the I frame. It is an instantaneous reduction in image quality.

Shorter GOPs seem to make the visual pulse less visible because the frequency goes up and it becomes visually similar to the random noise in the image.

 

[Nitsuj]

In a post somewhere Barry Green asked Tester13 if he could up the bit rate for that frame and Tester13 was positive about it. So hopefully that will work out. That is one of the things that has bugged me about the GH1 since I got it last year. Hopefully that can be addressed.

 

[wturber]

I saw pulsing from the start with 720p60. I don't recall if it was in the 1080 clips in the 60i wrapper, but I suspect it was. I doubt this has anything to do with being designed for NTSC. I think it has everything to do with the lack of B frames and a pretty crappy implementation of the AVCHD CODEC. Well, maybe "crappy" is too strong. Most of the people that see the footage from my GH1 oooooh and aaaaah over it. But then they just aren't looking at it as critically as I am. So for the consumer crowd, the CODEC is probably not "crappy" but is generally OK.

 

[cbrandin]

I think we've used the term "pulse" to describe two different things. I thought this thread started off being about the "pulse" seen in AVCHD Native 24p/25p mode, which seems to always appear right before an I frame as either a huge P frame or an extra I frame devoid of subsequent P frames, and is apparently a bug in the CODEC. The pulsing you're describing seems to be a normal result of a limited CODEC. Maybe I've misunderstood.

Chris