Here's the first grab. This was shot in Grand Central Terminal in NYC at about 10:00 a.m. There's daylight streaming through some windows up high and there's some tungsten lighting in there as well. It was overall kind of dark in there.
http://www.hdvinfo.net/media/bgreen/...XL2-GCT60i.jpg

This shot was in 60i mode, which should give the FX1 the most advantage it can have, and put the DVX and XL2 at the most disadvantage (they are both higher-res in progressive than in interlaced). Furthermore the DVX was hampered by having to be in squeeze mode, vs. the native 16:9 of the other two.

The DVX and FX1 had to be zoomed in a little bit to account for the not-as-wide-angle of the XL2. So the FX1 was wide-open aperture, as wide as we could get it, which at this zoom setting was f/1.8. The XL2 was also at f/1.8 and the DVX was at f/2.8. No gain was used.

This is an interlaced-mode grab so there are interlaced artifacts. This grab is taken from the DVD MPEG-2 file. No post-processing, no color correction, no gamma shift, no sharpening, no anything, not even studio-to-rgb correction. This is all three cameras taken directly from the firewire port (FX1 in full HDV mode), laid on the timeline, an identifying graphic added, and this is the output of the DVD MPEG-2 compression.

Now, from my observation point, I don't see where the FX1 is "blowing away" the other cameras... in fact, you can see much more detail in the shadows, and much more detail overall in the DVX and XL2 shots. And I think that squeeze mode acquits itself rather well as compared to the native 16:9 of the others... although the XL2 does have an advantage in detail as can be read in the words on the sign. The FX1 is dark because it just didn't have the sensitivity to properly render the scene, it was at wide-open aperture, as was the XL2. The DVX had to be stopped down 1.5 stops.

-----------------
This one is DVX/24P/Anamorphic/CineGamma, vs. FX1/HDV/CF24/CinemaTone Gamma, vs. XL2/24P/16:9/CineGamma.
http://www.hdvinfo.net/media/bgreen/...XL2-CP24an.jpg

Unfortunately the framing is not quite as good on the FX1, it's a little more zoomed-in, but still the differences are stark. The DVX is displaying the characteristic soft-contrast look of CineGamma, which could be "punched up" to deliver more contrast by lowering the master pedestal, something we didn't do. The FX1, on the other hand, is displaying the characteristic harsh-contrast/crushed-black look of CinemaTone Gamma. In this shot, for color, the XL2 is much greener and maybe a little sharper. The leaves on the tree were more accurately represented (color-wise) by the other cameras, but the grass and bushes look vibrant and cleaner on the XL2 shot. For detail the FX1 comes in last place. Look at the detail in the tree trunk (even though the FX1's zoomed in more!) Look at the lack of detail in the foreground bushes, the lack of detail in the splotchy grass at the base of the tree as compared to the cleanness of the XL2 or even the anamorphic softness of the DVX. For fine tree branch detail it's pretty close, maybe a slight edge to the XL2, and the FX1 unfortunately doesn't have quite the same framing so it's harder to see and judge. Much of the awfulness of the FX1's picture can be traced to the field-blending effects of CineFrame 24, but ... if you want filmlike footage, you're going to have to do something to it, whether CF24 or post-processing or something... again, for DVD footage, I think the DVX and XL2 are superior here. And if you like the contrast & deeper blacks of the FX1, you can always change that in the menu settings (we didn't have a monitor with us and couldn't a/b the shots).

------------------
This time we shot with each camera towards its best advantage, so the DVX is 24P/thin/anamorphic, the FX1 is HDV/60i, and the XL2 is 24P/thin/16:9.

http://www.hdvinfo.net/media/bgreen/...XL2-tunnel.jpg

Now, in this shot, I'm not seeing a whole lot of detail difference between them -- other than the harsher contrast of the FX1 vs. the softer contrast of DVX CineGamma, but the actual raw detail looks about the same. The XL2 was unfortunately overexposed by about a half-stop (the DVX was at f/6.8, the other cameras at f/4.0, which puts the DVX a full 1.5 stops faster than the FX1 even when in progressive mode!) The XL2 should probably have been at f/4.8, but without an A/B monitor it's hard to get all that completely accurate in the field.

Here the FX1 acquits itself well, but again, this is a still shot -- you have to keep in mind that while you're watching the DVD, the FX1 looks like a VIDEO camera (because of 60i) whereas the XL2 and FX1 in this shot look like movie cameras. Very different look.

(continued)

I guess in summary, for DVD they all look fine, and there's certainly no knockout blow for one over the others, but I do think the DVX provides the most information simply from the fact that it's 1.5 to 2.5 stops faster than the FX1 and 1 to 1.5 stops faster than the XL2. It sees more in the dark, and sees deeper into the shadows. And the FX1 loses resolution to get its film look, whereas the DVX and XL2 gain resolution to produce their film look. That's why I say, for DVD production, the DVX is the best with the XL2 very close second, and the FX1 just isn't as suitable (unless your goal is to make 60i DVD's under bright lighting conditions, in which case it'll look about as good as the other two). For HD interlaced production it's no contest, the FX1 delivers the goods. But for DVD release... I'm not seeing any advantage to the FX1 (other than a lower-noise signal), and I'd hate to give up real manual lens control, 7 gamma curves, 4 color matrices, real XLR audio support, true 24P, true SMPTE time code, and all the other things the DVX and XL2 do, just to get... what, a more expensive camera with worse low-light performance and lower latitude? Doesn't make sense for DVD production.

At least, that's my take, until someone proves otherwise.

Now, the one variable to discuss is downrez quality -- if I'm doing a lousy down-rez, then the whole test becomes invalid. To reiterate, I am using Vegas 5.0 for the downrez, for the following reasons:
1) I had it.
2) Douglas Spotted Eagle has said repeatedly that Vegas-downrezzed material looked much better than allowing the FX1 to downrez itself (convert HDV to firewire DV on the fly)
3) Doing the downrez in Vegas, vs. having the camera do it, let us avoid one compression cycle and two color-space conversions, which should benefit the FX1 footage
4) When set to "best" quality, Vegas uses Bicubic resampling, rather than "good" quality's bilinear (think I got that right...) so it should be a solid tool.

Even so, some question the validity of Vegas as the downrezzing tool of choice. So I tested Vegas against PhotoShop, just to see how it would compare. This should let us judge the quality of Vegas' resizing, independent of any other factor, at least in comparison to PhotoShop's resizing.

here's a small extraction blown up to 300% to compare:

http://www.hdvinfo.net/media/bgreen/PSvsV.JPG

To my eyes, Vegas did a pretty good job, way way better than Premiere 6.0 would have done. But the picture is a tiny bit softer with a little less contrast. All the important detail seems to be there though, (can't understand why the contrast would be softer... the black level looks lifted a little.) there's certainly not enough difference that one could point to Vegas and say "that's the problem, use a different re-sizer and the results would be night and day".

Furthermore, let's compare Vegas' resizing next to having the camera do the resizing. Here I have a shot where we shot a scene in HDV mode, and simultaneously captured the firewire output in DV Rack, so I have an HDV source and a camera-downrezzed DV frame. I then downrezzed the HDV source using Vegas, so we can compare how Vegas does vs. the camera itself:

http://www.hdvinfo.net/media/bgreen/...V-by-Vegas.jpg
http://www.hdvinfo.net/media/bgreen/...-by-camera.jpg

To my eyes they look pretty close, but I'd have to give the nod to Vegas. Look at the gray patch on the soccer ball in the middle of the shot -- Vegas is much more accurate in its representation, the FX1 smeared the detail in its shot (note, I'm looking at the original raw uncompression, whereas you'll be looking at JPG's, but it should be a valid comparison anyway). Overall the detail looks about the same, but there's some more mosquito noise on the FX1-downrez vs. the Vegas downrez, especially around the bells on the monkey's face.

(continued)

I don't know what to tell ya, it sure looks to me like Vegas is doing a good job of resizing to SD, better than the camera would itself and nearly as good as PhotoShop. I think the results of these tests are are fairly accurate, and downrezzed FX1 footage is, at best, comparable to (if not slightly softer than) native XL2 and DVX footage. There is no sharpness advantage that survives the downrez process.

To understand why the downrez process is not delivering more sharpness from the source, I went to a theoretical extreme to see if I could understand it and explain it.

Let's take a couple of givens: by the time the image is rendered in the frame, a DV image (or DVD image) will be a grid of pixels, 720 x 480. An HDV image will be a grid of pixels 1440 x 1080. For DVD release or for SD broadcast, one would have to down-rez that 1440 x 1080 down to 720 x 480. Some are arguing that the downrezzed image should be superior to the native image, but my testing is showing it not to be the case. Why?

Okay, think about it this way. Let's take the ultimate example of DV resolution. Let's pretend that you shot a checkerboard pattern, and you lined up and framed the shot so exactly perfectly that the frame in the camera ended up being a pixel-for-pixel checkerboard pattern. The frame in the camera would look like this:
http://www.hdvinfo.net/media/bgreen/DVCheckerboard.JPG

(note this represents a blowup of a small extraction of the full 720x480 frame). Every other pixel would be black, every other pixel would be white.

Now, let's assume we shot the identical same frame on the HDV camera. Seeing as the HDV camera has 1440 horizontal, or exactly twice as much horizontal pixels for the same perceived physical measurement in space, you would expect (and indeed get) an alternating series of two black pixels, then two white pixels. It'd look something like this:
http://www.hdvinfo.net/media/bgreen/HDVCheckerboard.JPG

(note we're being generous to the HDV camera for the sake of example here: because it has 1440 vs. 720 per line, it would indeed be exactly a 2:1 ratio on the horizontal. But because it's 1080 vertical vs. 480, it would *not* be exactly 2:1 on the vertical. Which means the downrez will be even less accurate in practice than in theory, but for our theory here we'll pretend it's 960 and call it "close enough").

Now, to downrez to standard-def, the HDV frame will have to be sampled and converted, from 1440 x 1080 (or 960 in our example) down to 720 x 480. What does it look like when you downrez 1440x960 down to 720x480?
http://www.hdvinfo.net/media/bgreen/...eckerboard.JPG

(That's the results from a PhotoShop downrez.) Looks pretty good, right? But... are the edges softer? Has there been a loss of detail? Let’s zoom in and see how crisp and sharp the edges are:
http://www.hdvinfo.net/media/bgreen/...zMagnified.JPG

They're not. There's rounding errors. There's averaging that goes on. There's approximation.

That's the way downrezzing works. You don't get the checkerboard, you get an approximation of pixel averaging, trying to keep all the data from all the pixels and making every pixel "count" towards the end goal. The only way you could get an accurate checkerboard pattern would be if your resize used a simple pixel decimation technique, throwing away 50% of the horizontal and 50% of the vertical pixels (which would only work for a strict checkerboard pattern, and wouldn't work so well on other types of images). Then you have to factor in that decimation wouldn't account for the 1080 to 480 conversion -- you *have* to have some pixel-averaging system going on. And when pixels are averaged together, detail is lost.

Let's put it to another type of test, taking into account the difference between 1080 and 960: I made a 1440x1080 grid, filled with a checkerboard pattern of pixels 2x2. Then downrezzed that to 720x480. Here's a blowup of what PhotoShop actually delivers, taking into account the difference in height that results because 480 is not half of 1080:
http://www.hdvinfo.net/media/bgreen/HDVRealDownrez.JPG

Now, the counter-argument goes, why does SDX900 footage look so much better on DVD than DVX footage? Obviously because the SDX, a much superior camera, can do a much more accurate representation to its frame than the DVX can. The real question would be, using the same glass, would a CineAlta do a better job on DVD than an SDX900 would? Would a $100,000 HDCAM camera deliver a better final image on DVD than a $25,000 camera? Using the same glass?

My assertion, based on this study and these theories, is that no, it would not. The SDX would make a better looking final image on DVD than the CineAlta would, because the SDX would use the same glass but would deliver a final 720x480 frame that would be more accurate than a downrezzed 1440x1080 frame from the CineAlta.

A CineAlta would probably make a better-looking DVD image than a DVX would, because the camera head is so much better that it would probably survive the downrez process better than native DVX. But the SDX is such a better camera head than the DVX that I propose that it would deliver a better-looking DVD image than downrezzed CineAlta.

And so it is with the DVX/FX1/XL2. The FX1 appears to be a comparable camera head, comparable lens etc. to the XL2/DVX... certainly in DV mode the FX1 is no better than the XL2 or DVX. So downrezzing FX1 footage does not hold up as well as DVX or XL2 footage.

That's my theory, anyway. Mouth is now wide open, my foot positioned right in front, waiting for to be shoved in by someone proving this theory wrong...

Certainly Vegas isn't the be-all, end-all of downsizing, and maybe some unsharp mask or some other technique would improve the FX1 footage (but couldn't we also do the same to the XL2 and DVX, negating the FX1's newfound advantage?)

For HD production the FX1 is the clear winner, as long as you like the interlaced look. But for making DVD's, I don't see the FX1 as offering any advantage, and in fact it demands sacrifices (in audio quality, in 24P mode, in image controls, in light sensitivity, in many ways).

When I get the FX1 in my hands again, I'd like to re-capture some tiny snippets of footage that I could then post so you guys could use the original raw data and make your own comparisons.

Awesome work Barry.

Still, your checkerboard tests did not take into account in the effect of DV color sampling versus HDV color sampling. Furthermore, your tests with what seems to be 0% and 100% luma don't address the bit depth advantage. Aquiring and editing in higher resolution at a given bit depth should --in theory-- render an image after resampling which could contain more resolution in dynamics. In other words, the result would be that of having --I don't really know but, say --an 18 bit ADC and 16 bit luma sampling instead of a 14 bit ADC and 8 bit luma. This idea would seem confirmed by your tests with the white teddy bear, where you compared the image from the cam's downsampling versus Vega's downsampling.

We do this sort of thing a lot in audio, when for example we use a 16 bit portable DAT to record a mono signal on both tracks, setting different input levels for each. When the two tracks are added back together in post the effective dynamic range of the system is greatly enhanced.

The question then is whether in practice, this enhanced luma dynamic resolution --and enhanced chroma spatial resolution-- play a big role in the quality of the final image. If they do, then, depending on the kind of content, there might be an edge in using the FX1/Z1 over other cams because of that factor. But then again, for that to really work the camera head would have to be very low-noise and have great lattitude. Even though noise seems to be quite low, lattitude does not appear to be until now one of the FX1/Z1's advantages. Then again most tests I have seen on lattitude were low-light shots, so the camera's noise reduction magic might have interfered by gating the shadows. And your tests with the cams in progressive scan and CF24 might have gone better if you could modify the FX1 gamma setting to something more similar to the DVX and XL2. I do not know if such a thing is possible.

It would be great if someone could run a checkerboard-like test similar to yours but with color patterns and gradients, and include DV and HDV compression, so as to be as real-life as possible.

Anyway, let's never forget Marshall Spight's wisdom: "The difference between theory and real life is that in theory, there is no difference between theory and real life, but in real life, there is a difference."

Very nice stuff, gentlemen.

And the pressure's on for me to get footage up. I didn't do so well last year, except for some jpegs.

heath

Hello Barry, I don't have the XL2 or DVX to compare the FX1 to but I do have the DSR300 and VX2000, in DV I don't think the XL2 nor DVX could be as good as the DSR300 for obvious reason and I'm sure you will agree with me, I also noticed that the FX1 ability to handle contrast is not all that good compare to my DSR300 and , and I also have seen some frame grab on this site which confirm what I noticed. I also agree that low light performance it not the best either, atleast 12 db below DSR300, BUT what I have seen in the final DVD is diffrent than what you saw, there is no DOUBT in my mind that footage shot on FX1 and delivered on DVD is much better than footage shot on DSR300, so footage shot with XL2 or DVX can not have better picture on final DVD since they are inferior to the bigger CCD and better electronics and lens of the DSR300. You mentioned soft picture with FX1 on final DVD, actually what you saw was a Cleaner picture, most often people think they see a sharper picture when there is a little grain or noise in it, and they think the picture without noise is soft but that is not so, I have looked at and compared footages for many hours before I decided to sell my 2 VX2000 and 2 DSR300, I'm no tech testing expert, I goes by what I see in the final product, but I think the FX1 deserve better that what you have been report on it.
BTW I have one DSR300 with an 18X Canon lens if anybody interest in it. (-:

Hi Khoi,

You are not the first one to report something like that. All I'm saying is, these are the tests I've run, these are the results I've gotten, and I don't know how else to do it. And I've listed the testing and methodology so you all can see where I went wrong, if I did.

So if someone wants to take the same footage and demonstrate a reasonable workflow (by which I mean, let's not go and get a $100,000 dedicated hardware resizer, let's keep it in context of the type of desktop video producers who are likely to be using this gear) then by all means I'd like to see it. I just have to re-capture small .m2t's, because all the ones I have now are 70 megabytes or more, and I can't trim them down without recompressing the footage... so I'll have to recapture them in like three-second chunks or something.

And, regarding Marshall Spight's quote:That's just brilliant. I think I'll be repeating that quite often, thanks for sharing it!

<<<-- Originally posted by Barry Green : Whoa -- let's hold up here. HDV has way more detail than DV. Way, way more. If you watch the original HDV source on a 1080i television, the difference is startling. -->>>

I of course meant when converted to dvd, but WHOA again, why hdv doesn't have more detail in dvd than dvd from dv?
Logically it should have like film transfers and other hd formats.

Strong opinions here, little proof.

How many line pairs can dvx, xl2 (pal&ntsc) or fx1 can really reproduce?
How does the amount of details changes with movement?

Lets take a theoretical case that we are using pal dv cameras and progressive display (all displays will be progressive in couple of years).
Fx1's resolution will be 1080x0.7(kell factor) = 756 and progressive pal dv would be 576, so the difference is 24%.

Of course progressive dv cam's resolution can't be that good, because otherwise they would produce interlace flicker with interlaced displays.
So there is lowpass filter, but where it cuts?

After that the difference should be more than 24% so why this doesn't show up with dvd tests?

Does the default sharpness setting cut the details out of fx1 or soft drawing lense? Or mpeg2 compression? Or something else?

<<<-- Originally posted by Jon Fordham : I am currently scheduled to do a shoot with the FX1 for Heath McKnight this week. As part of my initial tests, I plan on putting my Putora 7A9 chart in front of both the FX1 and the DVX100A. -->>>

Great, Jon!

Barry,

First off, if you go to www.womble.com and d/l the mpeg-vcr demo, it should allow you to cut up your clips for posting if required, leave the settings as is otherwise it will recompress the footage.

Second, the slight softening in the downconversion is entirely due to the the low pass filter effect of the interpolation. I believe I provided a valid technical description and cure, but if you are still unsure check out
http://kwon3d.com/theory/filtering/lpass.html
The bicubic spline is equivalent to a Butterworth with N=3 which reduces the frequency response of the band-pass near the scan line frequency, and that is why I suggest one with at least N=5 a fifth order interpolation to prevent the slight softening that you see. Alternatively an edge shaprpen will be just as good, and is not a cheat but merely serves to restore the loss of higher frequencies through the filter action of the interpolation. Only after this is done can the image comparison with DV source can be made in fairness.

For example here's how Photoshop describes the bicubic interpoltion method
'Bicubic (Smooth) for the slow but more precise method, resulting in the smoothest tonal gradations'

Third, I have thought of two more benefits ( the nine B's if you like) to downconversion from HDV>DVD so my list is now

B1- reduction of noise by > 50%
B2- removal of the halo effect
B3- reduction of artifacts > 50%
B4- improvement of any slight de-focusing faults
B5- improved colour space going into the mpeg compression for DVD
B6- choice of whether to have interlace or progressive DVD
B7- better framing from cropping if desired without resultion loss
B8- superior full resolution slow motion at 50% speed
B9- 2 stop (12db) improvement in low light response using a triple pixel readout post technique

I'm sure the discussion will continue as others post results

Quote: And I'm giving the FX1 the benefit of the doubt by making the DVD source from the original HDV footage, not a DV down-rez, because by using DV down-rez, you convert the original 4:2:0 MPEG-2 down to 4:1:1 DV, and then again to 4:2:0 MPEG-2 for the DVD authoring.

But Barry - no-one interested in best-possible quality would down-rez to a DV intermediate! Surely they could go to uncompressed, or one of the high-quality codecs like huffyuv or wavelet, using a high-quality rescaling algorithm and perhaps a touch of (re)sharpening as per John Jay's suggestions?

Having read this thread, I'm left wondering about the quality of the down-rez procedure in the MPEG encoder - could it be a factor in lack of sharpness issue?

Yes thanks Barry, good post. It has me thinking. And this is what I have come up with.
The Sony FX1 is a hard cam to compare with for this issue. As it shoots interlaced causing it's own problems in down conversion and also using pixel shift to enhance its resolution, so it is not the best example of a true SD to HD or Film downconversion comparison.
That said. Take some 720x480 FX1 DV footage, deinterlace, save as uncompressed. Then take the HDV footage (the same slow pan over some textured scene. Make it an outside shot with lots of distance. Close in shots will not show the difference)deinterlace, save as uncompressed as well. Colour correct the footage extensively, composite in a few objects/titles and FX or what have you. Then downrez. Check out the difference. Especially in the background.
Sharp lines, like your tile idea would be a good example if it was tilted on an angle. This would demonstrate the effective anti-aliasing and colour space increase of the down conversion. Hard lines such as stair railings with high contrast lighting would be a good example as well.
I was skeptical as well untill I saw a DVX progressive compared to a HD10 progressive. Both letterboxed 16:9. I then realized HDV did have a place in the 16:9 SD world.

Since I specifically want to use the FX1 to perform two functions as a wide camera view and possibly alternate camera view by pan/zoom/crop both to output in SD on DVD this issue of the best way to downrez is probably worth a thread of its own at some time. This will be a learning experience just like learning the most effective way to make compatible DVD's was in the last few years. IT might also differentiate between the various NLE's and their choices for downrez algorithms/parameters. Comments from David Newman would be welcome.

Ron Evans