The Stutteriness, the motion trail, the ghosting, oh the humanity!

[John Vincent]

[QUOTE=Joel AaronGrab an HVX200 and do the same test side by side at the same focal length.[/QUOTE]

I'd go to jail! I don't own one :)

john
evilgeniusentertainment.com

[Miklos Philips]

As someone was hinting at earlier the problem may be with the display refresh and not the camera. LCD displays have this stuttering problem.

[Carl Hicks]

Originally Posted by Miklos Philips

As someone was hinting at earlier the problem may be with the display refresh and not the camera. LCD displays have this stuttering problem.

There was much discussion about this in June on a thread. The problem turned out to be the monitor, not the camera. See this thread, post #15:

http://www.dvinfo.net/conf/showthrea...ght=LCD+trails

Regards, Carl

[Barry Green]

Doesn't make sense though. If you can see the double image on a still frame, how is that the monitor's fault?

Here's a grab from the mini35 footage that shows the motion trail.
http://www.icexpo.com/HD100/MotionTrails.jpg

That's what's in the .m2t, not some effect being generated by a monitor.

[Tim Dashwood]

Originally Posted by Barry Green

Doesn't make sense though. If you can see the double image on a still frame, how is that the monitor's fault?

Here's a grab from the mini35 footage that shows the motion trail.
http://www.icexpo.com/HD100/MotionTrails.jpg

That's what's in the .m2t, not some effect being generated by a monitor.

There is no doubt artifact trails exist and do show up in some situations. I wish Jaadgy would respond to his thread and post a sample.

In that other thread I had observed that the perception of a the trail changes depending on the display device - mostly due to refresh rate . My LCD connected to my computer clearly displays a trail with that sample. However, when watching the exact same file on a CRT HDTV studio monitor, it is inperceivable. It's very odd, and I'm still trying to figure out why.

[Steve Oakley]

the short answer is that LCD's have a slow refresh rate and are lagging behind the real image. cheaper and older LCD's have this problem, so if you are gonna use a LCD, get one with a fast refresh rate.

all that said, I have on tape the same motion artifacts shown in that still image. I guessing its a mpeg2 artifact of some sort from motion estimation failing.


Steve Oakley

[Stephan Ahonen]

Originally Posted by Barry Green

Here's a grab from the mini35 footage that shows the motion trail.
http://www.icexpo.com/HD100/MotionTrails.jpg

This is probably from bit starvation due to noise from high gain and the mini35's ground glass (noise eats up lots of bits) and fast motion (which isn't very good for codecs which rely on interframe similarities to keep bitrates down).

It's an extreme situation that would test the limits of any codec.

[Steve Mullen]

Originally Posted by Steve Oakley

... so if you are gonna use a LCD, get one with a fast refresh rate.

You mean fast response time.

And, ground glass, IF it is picked-up by the CCD -- it's not supposed to be which is why it spins -- is very fine moving detail with horrible (impossible) to compute motion vectors.

But this doesn't look like a motion trail as in a streak. It looks like exactly what the Motion Filter looks like. A nicely displaced image.

Hmmm. Is this a freeze frame from a FireWire capture? That's the only way to really see the actual data.

[Stephan Ahonen]

Originally Posted by Steve Mullen

But this doesn't look like a motion trail as in a streak. It looks like exactly what the Motion Filter looks like. A nicely displaced image.

I thought that too at first, my first draft of the response was "That's perfectly normal motion blur of a bright light source," but look a lot closer. You see a little bit of a "shadow" to the upper-right of each building. My theory is still that gain noise is causing bit starvation.

[Jaadgy Akanni]

I'm dying to post some examples of ghosting/strobing/motion trail from my HD100u but I must confess to you guys I still don't know how to do it. Somebody please give me some directions if it's not too much of an imposition. Thank you.

[Barry Green]

Originally Posted by Stephan Ahonen

You see a little bit of a "shadow" to the upper-right of each building. My theory is still that gain noise is causing bit starvation.

That's exactly what I thought it was too. It was my theory way back then that the mini35 pushed the HDV codec way too hard, and this was the result. It's most definitely in the footage, it's a firewire capture of the frame, and it happened all sorts of places -- on the model's hand as she brushed her hair, etc. It's a lagging multiple-image artifact throughout the footage, but -- curiously -- only the night footage. I didn't really find any evidence of it in the daytime stuff, only in the night stuff.

[Steve Mullen]

Originally Posted by Barry Green

I didn't really find any evidence of it in the daytime stuff, only in the night stuff.

Did the night footage have more noise?

Let's assume bit starvation. Why/HOW could that lead to a "perfect" lagging copy of a moving image, but not lead to any maco-blocking?

And, why would the copy be faint?

-----------------

When I was capturing 720p30 from the first JVC camcorders via component analog I found the same artifact. I was using a clapper. I noticed it because I was trying to see how the analog capture board maintained sync with audio captured directly from my Mac's audio input. I had a consistent second image in each frame. (At 30p there was one frame with 2 images. At 60p, the same frame--with 2 images--was repeated twice.)

I sent a frame grab to JVC Tokyo asking HOW this could possibly happen. They said they had no idea. And, I still have no idea. I even considered the CCD was sampling at 60i.

Much later, when JVC said they would have a Motion Smoothing function in the HD100 I published, in Video System, my explanation of how it could be done WITHIN the encoder using 60p information.

=============== Temporal oversampling

"Every 1/48th or 1/60th second a new frame enters the MPEG-2 encoder. Thus, the encoder is supplied motion video that is sampled at 48fps or 60fps. These rates are all high enough to avoid temporal aliasing.

Compared to the 24fps and 30fps that will be recorded to tape -- the input is “oversampled.” Oversampling is very often employed by analog-to-digital converters to improve the quality of audio and video. Oversampling provides samples that are far more accurate.

The question is “how” are the two samples converted to one sample. There are TWO obvious possibilities. One is that the two frames -- 1/60th second apart --are simply blended together before being fed to the MPEG-2 encoder. This is much like lowering the shutter-speed to increase blur.

Alternately, the MPEG-2 encoder identifies "objects" that are moving in ways that will yield judder, and applies blur to only these objects. Although this seems it might be very hard to accomplish, MPEG-2 encoders already identify objects within a scene and determine their Motion Vectors (movement paths).

They use these vectors to "move" objects in recreated frames during playback. So, it wouldn't be that hard for the encoder to identify vectors that are so large they will cause judder.

Once identified, an object would be blended with the same object in the previous frame to simulate motion blur. (Remember, the previous frame from 1/60th second earlier, is available.)"


<< It's also VERY possible motion vectors are not sensed in the encoder itself. A motion sensing circuit would compare frames looking (via correlation) for excessive motion. Once excessive motion between frames is found, this frame and the previous frame are blended to simulate motion blur. (Remember, the previous frame from 1/60th second earlier, is available.) >>>


Looking at this more closely, frames 1/30th apart are compared for the amount of motion correlation between them. If it's high, only these frames are encoded. But, if it's low, then the frame is blended with a half-intensity frame from 1/60th before. Note that the "previous" frame that's blended will have an image half-way between the 30p images. This would be our residual image.

Note also that it's not necessary to blend an image from 1/60th second before. The blend frame could also be the one 1/30th second before that's been held in a buffer. It's just that one would get smoother motion if one had a 60p CCD than a 30p CCD. Which means the first generation JVC could have had the same Smoothing Filter. The result would just not look as good as the HD100.

But, of course, the Motion Filter was turned OFF.

UNLESS, the filter is always active.

PERHAPS, the only thing the OFF setting does is raise the threshold for "blending."

But, why would the spinning filter exceed the higher threshold? Because the speckles in a spinning filter are NOT making random movements that are characteristics of "noise." They are tracing arcs. They are movement. Not noise. Hence they trigger the frame blending.

Now, of course, a correlation sensor is NOT perfect. To some degree it is also sensitive to noise. Which is why your night shots got blended. The noise pushed the motion level over the threshold. In the day, the lack of noise kept the threshold from being crossed.

If I'm right, then the way one focuses on the filter is critical. And, spinning might be different from vibrating.

[Barry Green]

And, spinning might be different from vibrating.

For clarification, in these shots the ground glass wasn't spinning, it was vibrating, this was the Series 400 Mini35.

We were quite certain to have motion smoothing turned off.

I have other examples that show the trails persisting. Stepping frame-by-frame through Vegas, it does seem like it's retaining a shadow of the prior frame. However, the trail persists over what appears to be a full GOP. On a locked-down tripod shot with a moving subject you can make out the residual effects of the motion trail for several frames, if not a full six frames. So it definitely seems to me to be an issue with the MPEG encoder, rather than something tied to the motion smoothing (although I guess it could be something to do with the motion smoothing creating the trail...)

But it doesn't seem like it. With motion smoothing you pretty much get two frames blended together, with one at less opacity. With this trail effect what's happening is the last edge of an object is staying on the screen, like a 1-pixel outline of where the object was. I'm watching the girl walking towards us at night, and her shoulder is moving through the frame left to right, and in each successive frame there's a thin outline of where her shoulder met the background from the previous frame. It's not like it's the whole previous frame blended into the new frame, it's just the edge... the transition point... that's what's saved and carried forward.

And, get this -- it seems more noticeable in out-of-focus shots! We did a shot where the girl walks towards us (and the guy walks past her, and she turns and looks back at him and then laughs as she turns back towards us) and she walks through the shot and off screen, and as she gets too close to the camera we quit racking focus so she goes very blurry. It's in those blurry segments where the motion trails are most easily seen. Which ties in with Chad Terpstra's initial tests.

Now, I'm not sure that it's the out-of-focusness that's causing the trails to show up more, or if it's something as simple as the fact that when she's closest to the camera, she's larger in the frame, which means the motion is more pronounced and so it makes it easier to see the trail...

Check this one out:
http://www.icexpo.com/HD100/MotionTrails2B.jpg

In that shot there's a massive trail behind her trailing leg, and her shoe is actually imaged three distinct times. That's not normal motion blur (which would have resulted in a continuous smear of the image), that's a distinct trailed image that's repeated so that there are three distinct elements. For the outline artifact behind her leg, look to the upper left, or download this picture and toggle back and forth between them (this is the frame right prior to that frame; looking at this frame and then toggling to that next frame you can see how the prior frame's trailing edge is persisting into the new frame... http://www.icexpo.com/HD100/MotionTrails2A.jpg)

Here's another sequence:
http://www.icexpo.com/HD100/MotionTrails3A.jpg
http://www.icexpo.com/HD100/MotionTrails3B.jpg
http://www.icexpo.com/HD100/MotionTrails3C.jpg
http://www.icexpo.com/HD100/MotionTrails3D.jpg
http://www.icexpo.com/HD100/MotionTrails3E.jpg

As you step through these frames you'll see huge red outlines where her arms were in the previous frames. Look at the trail on her left arm between 3B and 3C; 3C leaves a big red vertical ghosted column behind where her left arm was in 3B. And on her right arm you can see very pronounced trailing edge artifacts, on 3B it looks like I can make out at least three frames of ghosted imagery, three layers of where her hand used to be. 3D is even worse because all three of those layers are still there (although fainter) and a new layer is created from her hand's prior position in 3C. By the time we get to 3E the individual outlines are not so clearly defined anymore but there's still a large red "disturbance in the force" where the discrete outlines used to be.

Here's another sequence; look at her right hand and you can clearly see how the trails effect leaves just the outline of the edge of her hand, not her whole hand. Drop these frames onto the timeline in your NLE and step back & forth between them and it'll be very easy to see exactly what's happening.
http://www.icexpo.com/HD100/MotionTrails4A.jpg
http://www.icexpo.com/HD100/MotionTrails4B.jpg
http://www.icexpo.com/HD100/MotionTrails4C.jpg

EDIT: okay, looked through some more and I found one that's pretty decently in focus and yet still exhibits the large red trailing/ghosting effect, so I don't think focus really has much to do with it. Look at this shot, there's a large red trailing artifact under her left forearm and a two-layer ghosting outline under her left tricep. And this shot was in focus, so -- perhaps the out-of-focus thing is a red herring?
http://www.icexpo.com/HD100/MotionTrails5A.jpg

[Steve Mullen]

I'll look at your test in the morning. You mention the possibility it lasted a full GOP. In the DV tests, didn't you guys find that sometimes the JVC got stuck for a GOP? Did it leave a gap for a GOP or did freeze the I-frame for a GOP?

Your description sounds like the I-frame content keep being recreated--faintly -- in the next (up to) 5 frames.

I wonder if after all this time folks are finding strange encoder behavior?

I'll try to grab a frame my test tape I shot a long while back.

[Jaadgy Akanni]

Thank you Barry for posting those. That's exactly what i'm talking about.