Does Sony HDR-CX550V really captures HD(1920x1080) video? or it is interpolated?

A pixel contains color information and thus your x3 is erroneous.

CX550 does not capture 1080 progressive.
It is 1080 interlaced.

That's a different question.

Recognizing that I am a layman and may not express this perfectly or may be missing some aspect:

The post describes the action of a 3-chip camcorder where values for Red, Green, and Blue are captured by separate sensor chips with filters determining which color a given chip is measuring. So each chip must be capable of capturing the 1920 x 1080 total - around 2.1 MP. Having more pixels on each chip is better but the 2.1 MP is what each chip in a 3 chip camcorder must provide to get to a 1920 x 1080 picture. Having 3 chips of 2.1 MP each does not produce a 6.3 MP video output.

In the single chip camcorder, other approaches are used to produce the final color value for the "onscreen pixels". This chip can contain 2.1 MP and still create the 1920 x 1080 picture. The 3 chip camcorder may provide more accurate color measurement but it doesn't produce a higher resolution output if each of the 3 chips has the same sensor count as the chip in a single-chip camcorder.

So the Sony (a single-chip design) is definitely capturing more than enough pixels to create an HD 1920 x 1080 picture. A recent post in another thread described how all these camcorders do some interpolation to try to derive the best color value for the smaller number of pixels it is going to display. But I think overall you're looking at an averaging of information downwards, and not any kind of interpolation upwards to create more output pixels than were gathered at the sensors.

Thanks for the good explanation!

Well, I'm a bit nervous about how unsophisticated my explanation might be. I think my point that three chips measuring RGB colors don't produce an output pixel resolution at three times their additive resolution is clearly correct. But I just went back through Steve Mullen's guide to the XR12 and CX12 and there is some truly sophisticated stuff going on in all these camcorders, whether three chip like the Panny TM700 or one chip like most of the others in this price range.

Reading through other posts and Steve's guide, a few quick comments, and then I'll put some links in to helpful posts.

1. For primarily historical reasons, it looks like these camcorder sensors are typically 4:3 by design. So you can take a 4:3 aspect ratio picture at the full sensor's pixel count.

2. HD video is using the 16:9 aspect ratio, so the pixels for it are carved out of the bigger 4:3 rectangle and that video doesn't actually use all the pixels on the sensor.

3. Getting the "right color" measured involves taking clusters of adjacent light-capturing elements and running math on them in clusters to get an interpolated color for the group, which then gets applied to output pixels.

4. Start with a given sensor resolution - these clusters might only include four light-capturing elements, so that limits the accuracy of the color measurement.

5. Now double the sensor's light-capturing element, and each cluster used for averaging contains more data points and thus the accuracy of the interpolated measurement increases, improving the individual output pixels which of course improves what we perceive.

6. This goes on a for a while until diminishing returns kick in and there's not much point in making the sensor capture more data. The post below suggests that we're pretty much already there for many purposes in these camcorders, given a 1920 x 1080 output resolution. Ignoring progressive vs interlaced video for the moment, you only need so much data input before your quality doesn't get much better.

7. The initial HD camcorders had sensor chips that produced minimum color accuracy and detail because there just wasn't that much data to use for interpolation. They still captured enough data to produce true HD video but not enough to really produce great results.

8. The current crop of sensor chips provides many more samples and thus a much better HD picture. That's probably the best answer to the original question - is 4.x megapixels enough to produce real, usable HD in light of the RGB discussion: yes, it's enough and in fact quite good for that purpose. The RGB discussion made it seem like you needed 3 times the 1920 x 1080 pixel count, but that was because it didn't account for each sensor essentially being overlaid on the other two mathematically to get a color value.

Here's one posting that starts talking about this maybe halfway through:

http://www.avsforum.com/avs-vb/showt...2#post18519022

Here's another:

http://www.avsforum.com/avs-vb/showt...8#post18504778

You're on the right track, but your terminology is a bit off. An HD frame has 1920x1080 pixels. Each pixel is defined as a 24-bit triplet with red, green, and blue information in it.

Then there are sensor sites.. the actual photodiodes that measure light. They are inherently without color sensitivity, so they need to be "colored". One way is to filter on a per-pixel basis.. this is how DSLRs and other digital still cameras all work. A modern camera will have millions of pixels, and take 14-bit samples though a filter at each pixel site. The R and G components for a blue sensor site are, in fact, interpolated from the red and green neighboring pixels. When you have 10 or 20 million sensor sites, the slight change of the interpolation being off doesn't matter much. On consumer SD video cameras, with 1/3 megapixel, this interpolation can be obvious.

Another approach is to use a diachroic prisim to split the light into R, G, and B paths, and use three sensors.

In either case, if you don't have about 6 million sensor sites, there will be some interpolation, yes. This isn't always a bad thing... pro-level HD camcorders have been using interpolation for years. They still use three chips, but offset them, and use lower resolution sensors, which ups the size of the pixels and makes the camera more light sensitive. Color is going to get subsampled anyway (all consumer and prosumer formats do 4:1:1 or, more often, 4:2:0 color subsampling anyway)... so full color isn't inherently that useful. But interpolation can lead to color distortions that is visible, it depends on how its done.


Yup. The CX550 uses a 1/2.88" sensor with 6631K photo diode sites on it. This is one of these weird Sony sensors that's basically rotated 45 degrees. So they interpolate twice the number of actual pixels, but each interpolation is a little closer to being correct than the usual rectangular arrangement. Or something like that. They also use a non-Bayer color mask, which has even more green sensors than Bayer, and in some cases, maybe some unfiltered. This is called "ClearVid".

This is essentially the same overall idea that FujiFilm introduced as SuperCCD in their still cameras... only it's on CMOS, and for video. Anyway, you can read about ClearVid here: http://www.sony.co.uk/res/attachment...7477501978.pdf and here: http://www.imaging-resource.com/NEWS/1138169251.html

Like most sensors, this is a 4:3 image cropped to 16:9, so yeah, you only use 4150K pixels for this interpolation. Basically, they interpolate to 8Mpixel based on their 45-degree pattern being once again made rectangular, then they downscale to 2Mpixel.

So, as good and sharp as a three-chip camera with full 1920x1080 sensors? Unlikely. But they're doing the same as everyone else, trying to make intelligent trade-offs between light sensitivity, color quality, and cost. The proof is ultimately in the image.

As Tom Gull pointed out, 3-chips are for color accuracy not for increased sharpness or clarity.

3-chip is for color accuracy, yes. Full native resolution is for sharpness or clarity.. you cannot get any closer to 1920x1080 than starting out with 1920x1080 pixels.

If you have both, there's no need for any interpolation. Any camera still doing interpolation is highly likely to demonstrate less sharpness/clarity.

The pixels in excess of 2MP (1920x1080) are used for digital stabilization - moving a 1920x1080 frame within a larger window of pixels to reduce apparent jiggle - and for the higher resolution required of still photos.

Don't confuse the number of pixels with the number of bits/bytes of information in a frame or the number of transistors in the sensor(s). Thus, you only need 2 million pixels to render 1920x1080 frame, but you need about 24x that if you want 24-bit color depth (one 8-bit byte of intensity for each primary color) and much more than that for digital image stabilization capability.

Actually, that's not true anymore of most cameras.

For one, they nearly all use optical image stabilization, not digital. The optics are moving, and have a dramatically larger range than a few extra pixels. And this system is also fully function in still mode -- no extra pixels needed.

The reason we've always had spare pixels around has been the 16:9 format... the industry standard for imaging chips is 4:3. So to get a perfect 1920x1080, you need a chip with at least 1920x1440 pixels.

If you've looked around, other than Panasonic's 3-chippers, all of the leading edge consumer camcorders don't simply have a few extra pixels, but many extras. Sony's using 6 MPixel chips, Canon and Sanyo use 8Mpixel, JVC's up to 10Mpixel on some cameras. They're adding the additional pixels so they can use a single sensor and some form of pixel binning... multiple individual sensors per pixel to eliminate or reduce the effects of Bayer interpolation.

In a straight 1920x1080 sensor on a single chip, you usually have a Bayer color filter, which means you have only 1,036,800 green pixels, 518,400 red pixels, and 518,400 blue pixels. There's no need for spatial interpolation, but without color interpolation, you're going to have one funky looking' image. Thus, for every green pixels, you interpolate between red and blue neighbors on either side, to make an educated guess about that pixel's color, etc.


I believe you're the one who's confused... there's no way anyone needs 24x the number of pixels in a frame... and no serious camera is doing digital image stabilization anymore.

Each sensor site on a modern sensor produces 12-bits to 14-bits of luma. You would like at least three of these, one red, one green, one blue, per pixel for a full color image without interpolation. If you have that... a total of 6 million sensor sites, you can live without interpolation. If you have less, you need some interpolation still.

So let's look at a super high resolution imager, like on that aforementioned JVC camera. You can still have a Bayer pattern over your single CMOS sensor, which will look something like this:

R G R G R G R G R G
G B G B G B G B G B
R G R G R G R G R G
G B G B G B G B G B

For the video crop, you'll actually have 3840x2160 sensor sites. If you did the normal Bayer pattern interpolation on that, you'd get 3840x2160 pixels, just as on any old digital still camera.

But "bin" the sensor sites by four, and it's easy to see you get one R, one B, and two G's per bin. When "bin" = pixel, you now have 1920x1080 pixels, each with independent R, G, and B information, no need for interpolation.

But that's one big sensor... and some very small pixels. There are other ways. Here's what Sony does... remember Sony? The topic here?

Sony has a 6Mpixel sensor, but it's weird. Take that rectangular Bayer pattern I mentioned above and turn it 45 degrees. Now replace some of the R's and B's with extra G's, and you'll have what Sony calls "ClearVid" technology (same idea as FujiFilm's "SuperCCD"). Once you tilt the array, you find it's now intersitital relative to the horizontal... lines of pixels kind of zig-zagging from row to row.

So what Sony does it take four million of these and interpolate the whole array to an effective eight million color pixels. They're still doing color interpolation, in fact more of it, but the point of all this is that the distances for interpolated pixels, given the "diamond" pattern you get with the rotated sensor, is shorter than for a usual rectangular array. Thus more accurate... or at least that's their claim. Once they've interpolated out 8 million full color pixels, they downrez to 2 million.

So yes, there is interpolation in the Sony CX550.. that was the original. It's done in the interest of better color without the need for ultra tiny sensor sites... it's s compromise between a single 1920x1080 sensor, with the inherent color distortion, the 3840x2160 area needed for full color in a single sensor without interpolation, and the extra expense of going to a 3-chip system.

As for digital image stabilization... I have no idea if Sony's added that on top of their optical in these new models. They are doing very good stabilization these days. But bottom line... most of the pixels still go to the image. And for any decent DIS, you need pixels on the sides, too, not just top and bottom. Given their use of 4 million for the image out of a nominally 6 million pixel sensor, I claim they're doing nearly all of their stabilization optically... they don't have enough for digital stabilization. And that's good... I have a Sony camera from their days of digital-only stabilization, second one (HVR-A1), actually. It's a weak stabilization system... not enough spare pixels to be effective.

Here's an interview with some Sony people about how ClearVid works:http://www.camcorderinfo.com/content...-of-Choice.htm. Here's the ExMOR/ClearVid Whitepaper: http://www.sony.co.uk/res/attachment...7477501978.pdf

I'm 99% certain that the 3-Way Shake Canceling first introduced in the CX500 and CX520 series is digital processing all the way. There is in fact a notice somewhere that with Active stabilization on you will lose some picture around the edges looking at the LCD. That is, they warn you to frame carefully because the area captured is slightly smaller all around with Active IS on. Also note that this kicked in between two models that are otherwise thought to have identical optics, sensor chips, hardware except the media, and the like - the XR500/520 and the CX500/20. So unless you think they redid the optical stabilization between the two models in four months, the crowning touch of the stabilization improvements was digital. This matches cleanly with what you can see discussed on the Internet in various places from non-camcorder-vendors. Digital stabilization has taken great leaps recently and is much more effective than it used to be.

Traditionally in video cameras, that was true. 3 chip camcorders of years ago clearly outperformed single chippers in terms of color accuracy and saturation. However, it is not so simple any more. Single chippers can have excellent color saturation. All digital still cameras, whether point and shoot or DSLRs, are single chip devices, and the color accuracy and color control of a DSLR is excellent. There never was a need to make a 3-chip DSLR. My suspicion is that in the past, in-camera processors were fast enough to handle the color clean-up necessary for a single chip still camera to achieve excellent results, but the electronics were not fast enough to achieve the same quality for a single chip camcorder shooting video at 30 frames per second or 60 fields per second. Today, the in-camera processors have improved so much that they probably no longer face that problem.

Actually there are some significant advantages to have 3 chips/prism vs a singe chip even today. And the 3chips/prism have some significant design issues that require precision parts along with care assembly. But I agree is i not done or needed for color today. Or at best a slight advantage.

I think the reason we see them on HD digital video camera even today is that the relatively low resolution of the HD standard (2MP) and the small sensors make its reasonable size,cost and the parts/assembly "work" for video. Try to go much larger for the same advantages in a larger camera and you run into multiple issues. But it has been attempted, but never made it to market (see below). Today, the closest we have to in the digital still camera to the 3 chip/prism is the one chip/3 layer foveon design by Sigma.

But the holy grail of getting away from the Bayer singe chip still exist- if nothing for the ultimate PQ.

"Back in 2000, Hasselblad planned a medium format 16 MP DSLR with three
separate Foveon-made CMOS sensors, one each for red, blue and green.
It was to have been called the Hasselblad DFinity.

http://everything2.com/title/Hasselblad+DFinity

The project was never completed. Some prototypes were shown to a
select few journalists, and that was it.

Foveon went on to design and manufacture a three-layer sensor that is
used in Sigma DSLRs and Sigma P&S digicams"
.

Skoor, thanks for the additional info. Years ago, the color differences between single chip and three chip camcorders were dramatic - easily visible by even a casual observer. Nowadays the difference is slight or non-existent.

My guess is that among digital still cameras for the mass market, including DSLRs, there will never be three chip models. Perhaps there will be very high end models with three chips to sell to those with $$ and who can deal with the bulk, but the advances in in-camera color processing for single chip cameras will probably kill any real advantage three chip still cameras might have. Research and investment dollars flow to the mass market.

Right, when I bought my 3chip Sony TRV900, I was quite surprised how "bad" the single sensor video cam's colors looked in the 1999. It was very obvious. Today, looking at the Sony 550V, it is essentially the same as the Panasonic TM700. CCI actually shows the differences, but these are all really very good color performers. Heck, in the film days, film/lens had more coloration then I see in most high-end digital cameras. Bottom line, I would not make a cam decision for colors on the 3chip vs single chip design.

Today, it would be in the elimination of the blur filter to get a sharper image. But I assume the complexity of the 3 chip/mirror/alignment probably offset that advantage somewhat. The other would be Bayer issues the 3 chip avoids by having essentially the color filter logical (if not physical like the single chip foveon) in the same lens light path.

If does seem like the 3chip design might have a slight edge in low light assuming the prism light loss is negligible and the 3 sensors are physically bigger at the pixel level, which they should be for the 2mp HD standard. But look at the new pro AVCHD Panasonic video cam, a huge big sensor (from the GH1 family) and high resolution to boot. And no prism in sight nor would I ever expect to see one for the same reason we don't have the 3chip MF cameras- -cost, size, complexity vs real need for it today.

Well, it is called Optical SteadyShot w/ Active Mode.. that's got to be a pretty good clue. But you got me curious, and I always like to understand how these things work.

You're partly right... according to the published documentation, they're doing electronic roll stabilization (Canon and Panasonic systems don't address roll), which makes total sense. After all, in pure roll, you're rotating the camera along the lens axis. Optical stabilization can't help there. But the main X and Y correction is via a moving lens element, just as in any other OIS system.

That's also pretty smart, because while you run out of pixels very fast trying to compensate for X and Y movement, you don't with roll compensation. In fact, if you has a square imager, it would be technically possible to compensate against any possible roll, as long as the optical X and Y correction still worked.

They have a cute demo, in case anyone doesn't understand what this is doing: http://www.sony.jp/products/overseas...g_r/index.html


I think the timing of releases tells you very little about the timing of the actual design process, per se. But if you have the same lens and sensor, then no, they didn't likely redesign the optical parts. You could do the roll sensing entirely in software, but adding a sensor makes it dramatically faster. I suspect that's what they did... roll sensor plus new software, and you get one more dimensions of stabilization quickly.

That's also why they didn't have extra pixels built-in. In the bad old days of digital stabilization, you often had to choose between full stabilization or full video. I had a fairly high-end Sony Digital-8 camera, in the days of SD and tape, which was fairly clever about this. The sensor had extra pixels all around for digital stabilization. If you wanted 16:9, you could also use them for that... rather than the 16:9 cropping used on cheaper models. But it ate up the spare pixels, so no DIS.

The advantage has been fading. In the days of SD, it wasn't a question... you had 1/3 megapixel, and cheaper cameras usually not even that. Three 1/3" chips is always better than one 1/3" chip... if for no other reason than you get 3x the light per pixel, collectively. You might even go with 1/2 resolution sensors with offset, to boost light sensitivity at the cost of a little bit of interpolation.

But the single chip has to make a worse decision.. go to much smaller sensors and use multiple sensors per pixel, or keep the same sized sensors and deal with really, really visible interpolation artifacts.

When HD started, it was a similar choice. For the single chip, the interpolation wasn't so bad... same errors, but with 6x the resolution, they were 1/6th as visible (or more likely less than that, since the brain tends to threshold these things). The HD cameras still used the three chips with lower resolution and pixel offsets to deliver higher sensitivity.

In the latest round of consumer models, we see nearly every possible approach. Panasonic's using 3x3Mpixel sensors, for perfect 1920x1080 24-bit pixels. This approach actually sucked until they went to 1/4" sensors with modern sensitivity and low noise... the same thing helping the single-sensor cameras. JVC had models with 8Mpixels used for the video crop... that's one R, one B, and two G sensors per pixel... no interpolation needed either. Canon seems to have 6Mpixels in the video crop, which is pretty close to this, with a little interpolation I guess, but it allows larger sensor sites. And Sony's doing it with 4Mpixels with the ClearVid stuff already discussed.


That's the result of getting to one R, G, and B sensor per pixel, or close to it. Also larger pixels, and simply better sensor technology. Going to CMOS means more resolution (usually 14-bits per sensor, rather than 12-bits) and extremely effective noise management, for much less effective noise in the sensor.


Actually, Minolta did once... the Minolta RD-175. Far as I know, this was the only 3CCD DSLR. It was pretty bad. But the 3-chip arrangement made sense, back then it was 1.75Mpixels, just below HD resolution. You can see color errors in a scan-sized single-chip sensor HD camera... I've occasionally found such things in the video from my Sony HVR-A1.

Once you get much larger, though, color errors may technically be there, but they're invisible to the human eye. Thus, with consumer P&S cameras at 10Mpixels and DSLRs tending toward 20Mpixel+, there's absolutely no reason for this. Same choices that go into 4K cameras for digital cinema.


No, that's not it.. it's all about resolution. You never had DSLRs at the 1/3 Mpixel node.... those first retrofitted Nikons were over 1Mpixel. And they cost $20,000... and were only marketed to press photographers. No cares about color errors if you're shooting for monochrome newsprint.

By the modern era, there are just so many pixels, you don't see the errors. And any DSLR (or the new camcorders being based on DSLR-style large single sensors) in video mode will bin the pixels and downsize, so any color errors are completely invisible anyway. This main complaint is only the result of Bayer interpolation, and only something we care about if the resolution is low enough to actually see it.

Thanks for the clarification about why single chip video camcorders have progressed so far.

"3-Way Shake Canceling" is the roll stabilization and is what I was referring to as the "crowning touch" done digitally. It made a noticeable difference in my filming. I suspect this stabilization was just added into the "Active mode" menu item without being separately invoked.

The addition of a "roll sensor" hadn't even crossed my mind because algorithms for doing that EIS don't have to be told that roll occurred. However, I can see that a roll sensor might be useful in saying "check these frames" to speed the processing up as you suggest. Unless Sony tells us, we'll probably never know if the CX500s had a roll sensor added or not...

There's a whole thread of Panasonic users claiming that PQ difference is very noticeable. But that's not the same as saying the PQ from all the other camcorders this year is poor by any means. It means that 60p has an edge but I personally think it's going to be a slim edge.

One minor warning is that conversations about editing and playing 60p from this camcorder are very much like conversations about the highly-compressed AVCHD format three years ago. There's a lot of back and forth about having to buy powerful new computers to edit and play the video, and finding an editor that supports this 60p format. I just finished reading through the whole set of comments on the TM700 review at camcorderinfo.com. People there are mostly trading info as opposed to trying to sell someone on the Panny vs the Sony or otherwise. When I bought my first AVCHD camcorder, I worked on an older PC for about a month and then decided I really needed a new computer (a quad-core, which I bought in September of 2008, I think). So if cost is a factor and you don't have late-model hardware, you might want to read carefully to see if you'll need to buy more than just the camcorder.

It seems that the original-poster was really asking whether the CX500v image-sensor

(a) captures less than 1920x1080 'raw pixels', then upsamples the raw camera-signal to match the recording-format (1920x1080i)

(b) or, captures exactly 1920x1080 pixels, for a '1:1 pixel-mapping' (no resizing applied),

(c) or, captures more than 1920x1080 pixels, then downsamples the raw camera-signal to match the recording-format...

I think the previous posters did a fine job of answering this question.
A few things I'd like to add...

From a theoretical viewpoint, capturing more than 1920x1080 pixels may improve the final image-sharpness, since the pixel-oversampling can reduce the complexity/rolloff of the anti-aliasing filter. In audio-sampling, one reason (not the only one) for oversampling D/a and A/D converters is to reduce the complexity of the anti-aliasing filter. Without oversampling, the anti-aliasing filter must be designed for a sharp cutoff frequency, which is more difficult to design than a similar filter with a longer rolloff. I would guess the sample principle applies to optical systems.

Looking at the Video-capable DSLRs (Canon T2i, 7D, 5DM2, etc.), it seems there is such a thing as 'too many pixels.' The video-output from those units suffers from severe aliasing, which can manifest as a variety of visual artifacts (moiring is the most obvious.) That suggests the sensor, in movie-mode, decimates (i.e. subsamples) the pixel-sites to generate the video-signal. Doing so completely defeats the effect of the anti-aliasing filter, and hence, allows those moire artifacts into the final image.

Bear in mind, camcorder sensors are generally designed for video-capture first, so the designers are much less likely to use an overly high pixel-count which forces them to subsample/decimate in movie-mode. (Unless of course, camcorders start shipping with today's 15+ megapixel sensors to compete with pocket digicams...)

Correct me if I am wrong, but I suspect that the 37mm glass on the CX550 is not enough to resolve what would be considered to be true full HD resolution.

Because...? Are you suggesting that no camcorder can have full HD resolution with 37mm lenses? If so, what's the minimum size required?

I've never heard a suggestion before that the current lenses on any of the manufacturer's consumer camcorders are a physical limitation on the full HD capture. I'm not saying I can personally correct the posting if it's patently incorrect - maybe it's not. But I've never seen the question raised before and i"m wondering what the source for the thought might have been. You can post the Internet link if there's a discussion of this somewhere.