Pixel Shifting for dummies

J

Jockomo

Well-known member
Wildly oversimplified, and frankly it ignores so many real-world circumstances as to be functionally useless. I mean, I appreciate the effort he took in trying to explain the system, but the conclusions he leads you to are not justifiable in the real world.

First, he approaches it from the perspective that it would even be POSSIBLE to have a full-raster sensor instead of a pixel-shifted sensor. It ain't. Why do you think the pixel-shift system even came into play? Because it had to.

Here's the problem with that whole concept there: you cannot make a full-raster 1/3" CCD chip. It's not possible. Nobody can do it, nobody has done it, it would incinerate from heat if you tried to do it. Every manufacturer has had to try to come up with some sort of compromise. There's a limit of the # of pixels that can be read before the whole thing just overheats and goes up in flames. Which is why each manufacturer has had to compromise somewhere.

Canon chose to use interlaced chips. That cuts the speed of readout in half, it only has to read off half the pixels at a time. It also prevented them from doing 720p at all, and it's the reason they had to invent their "24F" recording mode -- because you cannot do 1920x1080 (or even 1440x1080) progressive CCDs in 1/3". It can't be done.

JVC tried to do 1280x720 @ 60P and found out that they couldn't. You cannot pack that many pixels on a chip and read them off at 60 frames a second; Dave Walton of JVC told me directly that the heat generated would have burned the chips up. Their solution was to split the chip in half; some people say that it's reading out different halves of the same chip, others have actually said that it's two separate CCDs physically butted together. That way they're only reading 360x720 at 60fps off the chips. They get progressive scan, they get 720/60p, but they also got the dreaded "split screen effect." And they don't have 1080/60i at all.

Sony took a middle ground, they used interlaced chips in the FX1/Z1U, and they also used spatial offset ("pixel shift"). They also couldn't do 720p and they couldn't do true 24p at all, they invented "cineframe 24".

Panasonic's design called for doing both 720p and 1080p and 1080i and variable frame rates. You cannot do that with an interlaced sensor, so they had to rule that out. That meant they had to use a progressive sensor, and we already know they can't even go 1280x720 on that sensor without the heat issues. They could maybe have done it with splitting the CCD in half, JVC-style, but JVC went through HELL for that decision. So Panasonic took the only route available to them that I can even think of: they lowered the pixel count on each chip to the threshhold where they could read the chips at a full 60P rate, and then used spatial offset ("pixel shift") to make up the missing resolution.

There was no other way to do it. The mythical 1920x1080 (or 1440x1080) progressive sensor doesn't exist and cannot exist. So which compromise do you want? Throw away progressive scan? Throw away true 24p? Throw away 720p? Throw away variable frame rates? None of those were an option, so they went with the option that made the most sense and allowed them to build the camcorder they had envisioned.

Every one of the approaches these manufacturers took involves compromise. There is no free lunch. And there's some 40,000 HVX200 owners out there who will testify that the Panasonic approach works just fine. All the cameras produce good images, of course, but there's only one 1/3" camcorder that does 1080 and 720 and variable frame rates, and the only way they could do that was with the spatial offset approach.

So now, flash forward a couple of years and there's a new option: CMOS. CMOS is an old technology that never worked very well as a sensor, but with recent advances in lithography and other developments they can now make CMOS chips that are largely competitive with CCDs. But CMOS has one big advantage that really means something for small-sensor sizes: CMOS is very low power draw, which means it's much lower heat. With CMOS they can get full-raster pixel density and still keep the heat under control. You can do full progressive scan with full raster chips. You can do 720 and 1080i. Except for one little nagging side effect: ROLLING SHUTTER. And all the problems that brings with it, from jello-vision to skew to scrolling bands through your video when shooting under magnetic ballasted lights, to big horizontal bands during flashes, and troubles with match-moving and compositing.

There is no free lunch, there will never be a free lunch, there's always going to be a tradeoff.

Three chips has always been better than one, except in the newest digital cameras that need to use a single sensor (because the lenses they use simply can't allow for the prism necessary for three chips). Why does Red use a single Bayer sensor instead of three chips (like every professional video camera on the market does?) Because you cannot use film lenses and a prism; it's not possible. So those cameras need to use a single sensor -- whether it's "better" or not doesn't enter into the equation, it's NOT POSSIBLE to use a three-chip system because it physically prevents the lens from working. So, those cameras use Bayer pattern sensors; the DSLRs and the Red use extremely high density to overcome the Bayer restrictions. Yet in that same article doppelbauer lambastes Bayer, yet many of the cameras producing some of the best images around are Bayer-pattern (speaking of top-notch DSLRs and the Red One). Bayer can work great, as long as you have way too many pixels on the chip to overcome the downsides of the demosaic process, but martin's article pretty much paints Bayer as the worst choice.

So: there's tradeoffs in EVERY engineering decision. It may be easy to draw some sketches that say that one or the other is a better solution, but the real world might disagree. Every camera manufacturer (except JVC) uses pixel shift because it works.
 
Last edited:
Thanks for putting some perspective to all of that Barry, interesting stuff.
After reading it (the linked page) I still had to shrug and say, but it's still CMOS and it's still intraframe.


I found the illustrations of how images get converted to pixels interesting though.
 
You must log in or register to reply here.
Back
Top