View Full Version : Why does 720 24p have less skew than 1080 24p
Thomas Horton June 23rd, 2009, 11:02 AM I just purchased this camera, and so far, I really like it.. I'm perplexed though.. 720 24p exhibits far less skew/wobble than 1080 24p. There is a huge difference between the two.. There is also less noise in 720 24p. I don't understand why there is such a difference in skew and wobble between the two 24p modes.. Each mode uses the entire sensor.
Thomas Horton June 23rd, 2009, 03:05 PM Forgot to mention.. Camera Model Number: Panasonic HPX300
Barry Green June 23rd, 2009, 11:30 PM 720/24p mode shows dramatically less skew than 1080/24p mode, that's true. But nobody has said as to why. If it was reading a smaller patch out of the sensor that might make sense, like a 1280x720 subset, but that isn't happening because the field of view doesn't change. So... we don't know.
However, this is my #1 recommendation for those who find themselves facing shots that the 1080/24p mode shows too much skew for their taste. Switching over to 720/24p for those particular shots can greatly reduce the skew, at the cost of some image sharpness (but still, it'll be sharper than an HVX/HPX170 in 720p mode, since AVC-Intra records the full raster). Just like on the HVX/HPX170, we would use 1080/24p for the main stuff, and swap over to 720/60p for slow-mo shots, and intermix them. For the 300 I'd use 1080/24p for anything and everything possible, but if the wobble/skew reared its head, swapping to 720/24p for those particular shots might result in softer but less-skewed imagery.
Thomas Horton June 24th, 2009, 04:59 PM Thanks Berry.. Great advice!
I can't help but think Panasonic could fix this with a firmware update...
They MUST!
Rick Bolton June 24th, 2009, 07:57 PM A WASG (wild ass scientific guess)
Might it have something to do with fewer cells being read & transmitted - allowing the captured image to be transmitted faster over the limited bandwidth of the CMOS?
The entire sensor is read (fov = constant) but only 70% (or whatever) are actually transmitted to file.
Flame protective suit remains close at hand :-)
Barry Green June 25th, 2009, 08:29 AM Well, that seems like a perfectly reasonable assumption, of course. But traditionally all the panasonics have worked in a 1920x1080 matrix for all internal processing, and then downconverted to the appropriate frame size for recording afterwards. So if they were scanning a smaller frame that would imply a significant departure from their prior methods.
If they were scanning 2/3 of the sensor, they would get a 1280x720 image off a 1920x1080 chip, but... I would certainly bet there'd be some funky aliasing or something going on if they tried that. Reading two pixels, then skipping one, then reading two more, then skipping one... seems like that'd cause some wild artifacts in the image, and I certainly never noticed any such artifacts in any of the 720p footage we shot!
Oooh, but what if they did some sort of bizarre inverse-pixel-shift scan of the chips, not scanning the same pixels off every chip identically, but maybe on the red chip they'd scan the first two out of three pixels, on the green they'd scan the first and third, and on the blue they'd scan the second and third... then they could do some demosaic process to rebuild a 720p image out of it while cutting down the chip scan time by 1/3 horizontally and by another 1/3 vertically, meaning an overall much faster scan... and simultaneously they'd have maintained a full 1920x1080 luminance grid; you typically get about 70% of a chip's resolution after a demosaic but that'd be exactly what they're looking for in a 720p image anyway, as 1280 = 67% of 1920 and 720 = 67% of 1080...
My programming days are behind me, so I don't know if this'd work or not, but it's the kind of wacky idea I'd try for fun back then. Seems like a way to get 100% full-res 720p from a 1920x1080 chipset while cutting the scan time of the chip to about half. Which would speed up the scan time and minimize skew...
Jonathan Levin June 26th, 2009, 10:43 AM This may not be the most brilliant question, but what is skew and what does it look like, and what causes it?
Jonathan
Barry Green June 26th, 2009, 10:50 AM Article here: CMOS Rolling Shutter (http://dvxuser.com/jason/CMOS-CCD/)
Jonathan Levin June 26th, 2009, 11:28 AM Barry,
Thanks so much for that. My Google search for skew came up short. This will help.
Jonathan
Jonathan Levin June 26th, 2009, 11:48 AM Read the article. Great info!
So do the Panasonics have rolling or global shutters? My impression seems that there are definite advantages with CMOS sensor technology, as does a camera with a global shutter.
Jonathan
David Heath June 26th, 2009, 05:18 PM Seems like a way to get 100% full-res 720p from a 1920x1080 chipset while cutting the scan time of the chip to about half. Which would speed up the scan time and minimize skew...
It's an interesting thought, but I'm not sure if it would be possible to read non-consecutive pixels easily off the chip. I'd assumed they were read off row by row, so no scope for speeding it up. (?) Which does leave the mystery of why 720 mode has less skew than 1080.....?
The best I can come up with as a theory assumes the limiting factor is not the max rate of reading off the chip, but the speed of signal processing, and an absence of large buffering between the chip and processor. Hence, the read rate off the chip is deliberately kept slow and even (in 1080 mode) in order not to overload the next stage with data. Switch to 720 mode and the signal processor has less data to deal with further down the chain, hence can process each frame in less time, hence the chips can be read faster without causing an overload. Any thoughts, Barry?
If the angle of view changed between 1080 and 720 modes, it would be obvious that a smaller (1280x720) area of the chip was being used, still with a pixel-pixel correlation between chip and output, and it would stand to reason that read out time would be far less. As the angle of view doesn't change, it seems that explanation can't be valid. Very strange.
Barry Green June 26th, 2009, 05:39 PM So do the Panasonics have rolling or global shutters?
All video cameras currently in common usage, that use CMOS, use rolling shutters.
My impression seems that there are definite advantages with CMOS sensor technology
There are definite advantages to the CMOS approach, such as being able to window the sensor for smaller frames at faster readout speeds, and individual pixel addressibility, and much lower heat and power use. The heat/power issue is really why you're seeing CMOS appear in small-chip video cameras; you can't make a 1920x1080 1/3" CCD chip, but Panasonic's already got a 1920x1080 1/3" CMOS chip and in August they'll have a 1920x1080 1/4" CMOS chip in the HMC40.
as does a camera with a global shutter.
Global is pretty much the holy grail of getting a CMOS camcorder with no annoying CMOS artifacts. However, a global shutter (while possible today) is a lot more complex engineering, and results in a significant cost increase and a loss of sensitivity and noise performance. We'll probably be rolling for a while.
Barry Green June 26th, 2009, 05:46 PM Hence, the read rate off the chip is deliberately kept slow and even (in 1080 mode) in order not to overload the next stage with data. Switch to 720 mode and the signal processor has less data to deal with further down the chain, hence can process each frame in less time, hence the chips can be read faster without causing an overload.
But it's the scan rate down the chip that is the determining factor in skew. So either they're just plain reading it faster in 720, and purposely just *not* in 1080, or they're reading less of the chip (which I can only assume means some manner of pixel-skipping, because the angle of view doesn't change). But I don't think that can be true, because after the initial complaints about the skew, if it was a simple case of speeding up the read/refresh rate in 1080/24p mode I am pretty sure they would have. So I think something else must be going on.
The Canon 5D Mk II is widely believed to be doing some manner of pixel-skipping, which (while I don't know whether it's true or not) it does make some sense because that's a big ol' massive CMOS sensor and there's pretty much no way they could read the entire thing a full 30x per second, so if they were pixel-skipping it'd reduce the burden.
CMOS chips should be individually pixel-addressable, so skipping pixels shouldn't be an issue, right? I don't know if the Canon does that, but everyone else seems to be convinced they do, so -- if they do, then the HPX300 should be able to as well.
But, yes, it seems puzzling. I've thrown a note on this over to an engineer friend of mine, asking about the reasonability of such an idea and looking for other hints as to what may be happening; if I hear back anything I'll post it.
Jan Crittenden Livingston June 28th, 2009, 07:26 PM Hi,
The difference between 1080 and 720 is the difference in the size of the raster. So that said, the engineers felt that if they left the 24p capture to reflect the true 24p motion blur(desirable) that since you would be using and moving the camera in accordance with motion judder consideration, they felt this would allow the best picture. They did study it, but, with loss of resolution and latitude they could resolve it, but who would vote for that. Why not shoot with aqn understaning of how the tool works.
Take a look at this video: Working with 3MOS Imagers on Vimeo (http://www.vimeo.com/5348383)
It was all shot at 1080p/24PN AVC-Intra.
Best,
Jan
Thomas Horton June 28th, 2009, 08:28 PM Jan,
I appreciate the input, but can you tell us why the HPX300 has far less skew in 720 24p than 1080 24p? The full CMOS chips (1920x1080 raster) are being utilized in both, but the only difference is that 720 24p is recorded @ 1280x720 (down rezed from 1920x1080)... Don't get me wrong, I love the camera, but just don't understand why there is such a difference in skew between the two 24p recording modes...
Thanks again Jan
Jan Crittenden Livingston June 30th, 2009, 07:52 PM Jan,
I appreciate the input, but can you tell us why the HPX300 has far less skew in 720 24p than 1080 24p? The full CMOS chips (1920x1080 raster) are being utilized in both, but the only difference is that 720 24p is recorded @ 1280x720 (down rezed from 1920x1080)... Don't get me wrong, I love the camera, but just don't understand why there is such a difference in skew between the two 24p recording modes...
Figure the diff between the two sizes, 1920 X1080 vs 1280 X 720. One is twice as large as the other. The smaller is clocking faster because it is smaller. Watch the video for ideas on how to work with it.
Best,
jan
Thomas Horton July 1st, 2009, 01:54 PM But... they are both being read off of the imagers (CMOS chips) at exactly the same full raster size.. 1920x1080. 720 24p mode is simply down rezzing after the fact.. after it has been processed from the imager.. If it was a case of smaller versus larger, then there would be a crop factor negative size value offset. Both 720 24p and 1080 24p clock the same... 24 fps..
I did watch the video.. and yes.. it is very helpful :)
Jan Crittenden Livingston July 1st, 2009, 06:00 PM But... they are both being read off of the imagers (CMOS chips) at exactly the same full raster size.. 1920x1080. 720 24p mode is simply down rezzing after the fact.. after it has been processed from the imager
I did watch the video.. and yes.. it is very helpful :)
But when it comes off the dsp, it is going to a smaller format. Its smaller, therefore less. Think about it, this all happens in time frame that I can't even measure, I am not an engineer.
Don't know that that helps, but....
Jan
David Heath July 11th, 2009, 03:34 PM But when it comes off the dsp, it is going to a smaller format. Its smaller, therefore less. Think about it, ...........
I can't see how that explains it, not if the 720p/24 image is derived from a full-sensor 1080p/24 image - and it must be so, since if it was as simple as using the middle half of the sensor (in area terms) the angle of view would change between 720 and 1080 modes for the same lens focal length. (And it doesn't.)
If we imagine the camera panned across a vertical line, lets assume the panning speed is such that in the 1080 frame, the bottom of the line is offset 60 pixels with respect to the top of frame. The skew then becomes 60/1920=1/32 of the frame width from top to bottom.
Do a downconvert to 720p, and those 60 pixels must scale down to 40 pixels (60x 1280/1920), so yes indeed, smaller, therefore less. BUT those 40 pixels still represent 1/32 (40/1280) of the frame width - and consequently exactly the same skew!!
|
|