Devil's Advocate

[Glenn Chan]

Ok this is my first post on this forum and I'm going to talk some trash about this camera. :D

My subjective judgement:
Something looks bizarre with the colors... flesh tones look wrong, color reproduction looks wrong.
The overall look doesn't look like 35mm, but I assume this is because the shots were un-graded.
As far as the overall look... it doesn't look that natural and doesn't have the pop that other cameras (film, video) are capable of (with color grading).

Technical:
A- Grayscale Tracking

If you look at the CML party NAB 06 shot:
http://www.siliconimaging.com/Digita...y_footage.html
Take a look at the DSC Labs chart the models are holding.

In the black and white chips in the middle of the chart, the grayscale tracking is off. Bumping saturation will make this easier to see. The chart is supposed to be black and white, but the recorded image isn't. The chart is still off even after taking (mis-)white balance into account. The tendency is red in the shadows.

B- Colors from the DSC chart lining up:
If you mask the chart and see the image on a vectorscope, you see that the colors aren't close to the DSC targets.
One of the flesh tone chips (the bottom right one) is distinctly off in hue and saturation.
Red is oversaturated... other colors are undersaturated.
Yellow is too orange-y.
Etc...

*That being said, this assumes that you think that the DSC colors should be right, and to some degree that the test was fair (i.e. metamerism index of lighting is low).

A true test would be to measure the spectral response of the camera. I don't know any camera manufacturers which publish such data... which is unfortunate if you're interested in color accuracy.

C- The Adobe RAW algorithm makes the image look terrible... there is very noticeable aliasing on the test chart if you zoom 200%.

D- On the green screen footage, the image looks soft. Perhaps this is because of a different de-bayer algorithm... or that the camera was shaky (doesn't look like shakiness though?).


2- Extended dynamic range capture:
In most of the stuff I've seen, a ""linear"" conversion of the imagery (i.e. linear as in gain down in the camera) results in a flat image lacking in (perceived) contrast and saturation. When shooting this way, the image really needs color grading. A print film look (i.e. s-shaped transfer function / s-shaped RGB curve) would be somewhat reasonable common denominator.

I'm curious what the SI camera is doing in this regard... what transfer function is being used, and what algorithms are being used so that image doesn't look "flat".

The attached images are:
A- The original image is "linearized" output from a F900. "Linear" in this sense means the gain on the camera was (effectively) turned down via the camera's LUT.
This image courtesy Steve Shaw of digitalpraxis.net
B- The image with more color is in was processed by color enhancement algorithms (contrast, saturation boost) I am working on.

They show what's possible with extended dynamic range capture. It's not that great an example... other cameras can get more latitude, or are at a lower price point. The Andromeda mod for the DVX100 gives very similar results (too bad about their workflow).

[Jason Rodriguez]

Please keep in mind that the files you're seeing are uncorrected dailies from Spoon or other sources . . . and some files are from very early beta cameras. So we have been tweaking things as time goes on.

Also the CML test was NOT a scientific test AND all the files except for the DNG file have been graded.

A - That was our wide dynamic range curve . . . it's a custom gamma curve, and naturally, if you're going to get a wide dynamic range image that's viewable on a gamma-corrected display, you're going to get "incorrect" greyscale tracking on a linear chart. If we took our linear image and simply applied a 2.2 gamma curve to it, only the upper 6-7 f-stops of the useable 10-11 f-stops of the linear signal would be visible on a display. As a result, you have to make a gamma curve that compensates for this effect, which will give a "flatter" image, but will also cause a normal perceptually linear chart to have non-linear characteristics. That's just the name of the game. If you download a chart from the Arri D-20 or any other wide-dynamic range camera, you will seem the same effect. Additionally, wide dynamic range film will behave the same way. The darks in the CML image go a little bit red because I wanted the image looking warmer rather than cooler (on the graded images-and please note the grading included curves manipulation, so the linearity of the file is completely lost at this point). Call that the artistic decision of the shooter :)

B - You can redo the color matrix with our RAW files any way you want. If you don't like how the colors are calibrated in the camera, change them in post. That's one of the beauties of RAW - we don't "bake" the color information into the image, so color-correction and manipulation becomes very flexible.

One thing to also keep in mind is that image was very dark/low-key . . . not exactly optimal lighting conditions to critically evalute the color rendition of a DSC chart-i.e., without the correct amount of light, they will NEVER hit the correct vectorscope marks. That wasn't the point of the shoot. The point was to use the chart as a reference with the other cameras that were shooting that evening. Also please keep in mind that no camera on the market in it's default configuration, or any digital still camera that thousands of photographer's laud over line up with what the DSC charts specifiy as "correct". Basically every camera on the market shoots "incorrect" color, and that's why people are always claiming their camera's color is "better" than the next manufacturer, or why they like Panasonic "color", or Sony "color", or Canon "color", etc.

The nice thing about us is that we shoot RAW, so you can completely remove the color matrix (which is what the other manufacturers are using to get their specific color), and apply any other matrix that you like. Shoot your own charts and come up with your own colorimetry for the camera.

C - The CML image was shot with a beta version of the CineformRAW codec . . . the aliasing you're noting has been fixed in the current versions of the codec.

D - The greenscreen shot is in motion . . . it's a jib-arm shot. As a result the focus drifted slightly and you're seeing motion blur from the camera. This will prevent the shot from looking "tack-sharp". There's also no edge-enhancement applied to the image, since applying edge-enhancement to a green-screen shot before the keying is done is a major "no-no" and will seriously degrade the ability to perform a proper key (since edge enhancement creates false edges around the subject).

Glad you brought up these issues, especially if others are asking this in the back of their heads :)

Thanks Glenn

[Jason Rodriguez]

BTW, those "linear" modes on the F900 are not linear at all, they are gamma corrected . . . logarithmic files are actually very similar to more extreme gamma correction (beyond 2.2), so if you're looking at a gamma corrected file, you're actually seeing something more akin to a logarithmic file than you are a true linear file.

The point of the gamma correction is to make an image appear "perceptually linear", which is a very important distinction to make.

So if you're fudging around a gamma corrected image (what you term "linear") to make it appear wide-dynamic range, you're basically taking the same total dynamic range of the chip and fudging it around with some modifications to the default gamma curve of the camera (which is what knee, black stretch, etc. are in those Sony cameras . . . they are just moving around the values on the default gamma curve of the camera to affect how the overall dynamic range of the chip is distributed in the visible domain). So you're actually just going from one style of gamma correction to another between your "linear" image and your "wide-dynamic range" image-they're all in the end gamma-corrected images, and as such in order to get a wide-dynamic range image visually, a true linear greyscale can't track linear visually-there will have to be some nonlinearity to the values in the greyscale.

[Glenn Chan]

To clarify:

By grayscale tracking, I am talking about whether a achromatic/colorless chart has color casts when recorded.

From your comments, it seemed that the chart didn't track because the grading was introducing color casts.

BTW, those "linear" modes on the F900 are not linear at all, they are gamma corrected . . . logarithmic files are actually very similar to more extreme gamma correction (beyond 2.2), so if you're looking at a gamma corrected file, you're actually seeing something more akin to a logarithmic file than you are a true linear file.

"Linear" is a terribly ambiguous term now, but linear in that case meant "not log"... i.e. something like
f(x) = constant * x ^ (1/2.2)
*I don't know the exact transfer function.
Changing the constant in that case is one way of increasing the dynamic range captured. The Panasonic HVX200, Varicam might be doing something like that.

The transfer function I assume was generated using Steve Shaw's "Digital Praxis LUT Builder & Sony CVP Gamma Builder"... my understanding is that it loads a custom LUT into the camera. He now has a log curve for the F900, hence linear meant not log.

**My use of terminology is terrible.

Basically every camera on the market shoots "incorrect" color, and that's why people are always claiming their camera's color is "better" than the next manufacturer, or why they like Panasonic "color", or Sony "color", or Canon "color", etc.

Well a very good / "correct" colorimetry would be something that matches the CIE color matching functions. Sony's RGB+E sensor design in their digital still camera F828 line does this to get better color accuracy (I don't think it gets "perfect" color accuracy though). Truelight's calibration device I believe uses four color filters to match the CMFs and supposedly does a very good job.

In a practical camera you may not want this since it hurts sensitivity (and has R&D costs), but getting as close as possible to the CIE CMFs would be ideal. The CIE CMFs would be the correct color... not whatever Canon/Panasonic/etc. is doing with their video cameras.
**The Panavision Genesis may be taking an alternative route, matching their camera's spectral response to be more film-like.

2- It might be interesting to perform a test like the following comparison tests at CML:
http://www.cinematography.net/digital-latitude.htm

Things to note:
A- Cameras seem to lose grayscale tracking (i.e. pick up color casts) in the extreme upper end of exposure. It looks like the Arri D20 compensates for this (a good thing IMO).
B- There is controversy over the F950 since it wasn't optimize for maximum exposure latitude. The F950 as tested was with the "standard" settings.

[Jason Rodriguez]

Yes, grading and white-balance adjustments, along with the desire for a "warmer" image have introduced slight color casts.

We created our own curves based on what we saw as optimum for the sensor to give wide dynamic range images or to maximize signal to noise ratio, while keeping noise at a minimum. We are not using any transfer function from Steve Shaw (who's tools are very nice I might add), nor are we just cranking up the exponent on the gamma curve, since that would give you a very odd image with a lot of noise in the shadows (gamma curves are VERY steep at the beginning).

For a while we played around with using the Thomson Filmstream LUT's, but soon realized that requiring a separate view LUT just to get your footage to look "normal" is not condusive to a great workflow, especially one for independent film-makers.

We provide the true photometrically linear data from the camera head if one desires to use that, and that provides you all the dynamic range the camera can produce. You can do a lin/log conversion, etc., whatever your hearts desires. But basically our 12-10 bit LUT's are made to give the end-user the full dynamic range of the camera, but in a 10-bit gamma corrected space rather than 12-bit linear, and the don't need another look-up curve to make the data seem "normal". The wider dynamic range LUT's of course will look a bit more washed out than a normal SMPTE curve, but they will not look like a LOG curve from a Cineon or the Filmstream curve, as those require reverse look-ups to look correct, and we didn't want our end users having to go through that extra step, especially since it really doesn't gain the user any more "useable" dynamic range (they typically end up exposing noise once you zero out the blacks and add contrast back to the image, and that's not what I would call "useable". For instance, try underexposing a Viper by a couple f-stops in Filmstream mode . . . you'll be getting detail in the washed out LOG image, but not detail that you can use, because once you try to add "pop" or correct contrast back to the image, you end up also exposing all the noise at that level as well, and that doesn't look too good).

I have seen the results on the CML website (I'm an active member there), and we will be posting tests like those on our website.

[Pete Tomov]

Any chance of getting the shapnest of the Dalsa/D20 with the SI? Maybe with a decent prime and uncompressed recording?
The softness of the F950/Viper seems more like a lens issue. I've never liked DigiPrimes.

P.S. The D20 kicks ass, I just wish I could afford to have it as a family camcorder.

[Jason Rodriguez]

Hi Pete,

Good question.

I think the main problem with trying to achieve the "sharpness" of those cameras is that with their high-resolution sensors, they're able to oversample to get HD resolutions.

I know Dalsa is a 2x oversample, the D-20 a 1.5x oversample. By oversampling using a pretty good algorithm, they're able to get more resolution into the frame that we can, that's just a simple matter of physics. It's like making an standard definition DVD from a Hi-Def source . . . even though both are standard definition, the Hi-Def version looks much "sharper".

Of course if you use a very simple algorithm for oversampling, then the resolution benefits are mitigated . . . in fact you may incur aliasing artifacts and moire from the oversampling process rather than the opposite. You have to apply some type of low-pass filtering when you oversample, if not, you're going to get problems, and then oversampling becomes a problem rather than a solution. The best example of this is when you take a sensor and remove the optical low-pass filter . . . you may get more resolution in the frame, but when the pattern of the object you're shooting goes beyond the Nyquist freq of the sensor, and you're getting objects that don't cover an entire pixel, you get nasty aliasing and moire patterns . . . more resolution now becomes a detriment rather than an advantage.

Suffice to say, both Arri and Dalsa are doing low-pass filtering in their oversampling algorithms to get HD resolutions . . . actually Dalsa doesn't do HD resolutions at all, it only does 4K, and then you might use something like a bicubic transform, etc. to get it to HD resolution. So their images look really nice and clean.

I wish I could say we could match them pixel-for-pixel, but the truth is it's probably physically impossible . . . of course there's only so much information you can pack into a HD frame (even if you're oversampling), so I think you would have to look closely to see the differences . . . i.e., it won't be a day-and-night difference between us and them, but there will be differences.

As you've nicely noted though, you can't own either of those cameras . . . ;)

[Pete Tomov]

Well the Viper and the F950 don't show much more resolution, at least with these lenses, so they're pretty much out of your way.

The Dalsa show horrible noise after heavy color correction where the SI actually hold up better but show something that apears to be compression artifacts(again, it's really heavy color correction).

I can't wait to try 12bit uncompressed recording, I dont' care how good this algorithm is, it's still 5 times less information.

Edit: I thought that since the Dalsa images were processed by...well Dalsa, I might try proccesing the other images as well.

Apparently the F950 simply doesn't have this resolution but from the Viper images I got sharpess that's really close to the Dalsa.Some small artifacts popped up but they're really hard to see and they look like crappy debayer algorithm.

[Jason Rodriguez]

BTW, Peter, which images are you grading from our camera?

We've made some adjustments to the codec to eliminate certain noise-related artifacts that were cropping up.

Also, I've done both Cineform RAW and Uncompressed, and for everything but green-screening, you really can't tell the difference when you're working with the footage.

Additionally, remember that 12-bit uncompressed doesn't edit in real-time. We record and output to a proprietary format and then re-wrap to DNG sequences which will require importing into After Effects, or pass through a RAW conversion program, and then process those images into a movie file. That can take some time. For instance, at 1 second per frame for rendering (demosaicing), that will take 24 hours to process on shot-hour of footage. The nice thing about DNG's though is that it's an "open" RAW format, meaning there is no licensing required and the spec is completely open for use by anyone.

CineForm RAW on the other hand is direct-to-disk and imports directly into Premiere Pro. You can mix/match it with any other footage in real-time, apply real-time effects, edit in real-time, and then output to tape, DVD, etc. Additionally you can playback the AVI files or QT files in any media player application, again, all in real-time.

Complaining about the quality of Cineform RAW is like complaining about HDCAM-SR. And Cineform RAW decodes the RAW bayer to 4:4:4 RGB, so you get really high-quality footage from the format. But we offer both compression and uncompressed because there are many uses for uncompressed . . . but shooting an entire film uncompressed is complete overkill IMHO.

[Pete Tomov]

I'm particularly talking about the greenscreen footage since all the other footage is recompressed.Green pixels on the out of focus green screen seem 'grouped' like in a DivX type compression.

I understand the issues of recording uncompressed but I'm sitting on a lot of CPU power anyways(I've got a AMD Athlon64 3800+ and and Dual Core 4800+ in my bedroom and I have my hd-sdi recording machine that's also AMD Athlon64 Dual Core 4800+) and I love to buy computers because I'm also sort of a 3D 'artist'. Anyway, I just wanted to test it.

Other than that I love what I've seen.Here's a grade test:

http://img109.imageshack.us/img109/9817/benchcc5pp.jpg

I was trying to get a saturated look since the original was flat.It turned out pretty well considering it was a WMV.Looks very Johnny English.

[Jason Rodriguez]

Your graded image looks nice, even for WM9 :)

BTW, there is absolutely no compression on the green-screen shots we have up . . . so I'm not sure what these "grouped" pixels are, but they surely can't be compression artifacts.

Can you take a screen-grab of what you're talking about, because I'm not seeing any artifacts in the green-screen images. They look pretty good to me.

Thanks.

[Pete Tomov]

Erm...I never saved that file but I tried to grade it again and...now it looks more like noise.I was awake for 34 hours when I did this that must've been it.

Sorry.

It's weird, when I increase the saturation I get really strong noise that I don't get on any of the other shots.

[Jason Rodriguez]

Hmm . . . that shot is already "graded" at least concerning saturation . . . I think the only grading that needs to be done is remove the green-screen spill which is affecting his skin tones . . . I'm not sure that the saturation needs to be bumped up any further for that greenscreen shot. Cranking saturation too much will always lead to more noise, that's just the nature of the math . . . especially if you're using the Adobe saturation tools . . . those are no good. I would recommend of color grading using either Discreet Combustion or Color Finesse.

You can post a sample of the color corrections that you're doing if you want to show me any odd noise areas you might be concerned about.

Thanks.

[Pete Tomov]

...so I checked out Combustion 4 and you're right, the image is incredibly clean.Now I just need to get used to the color curves control.

[Heath McKnight]

If I'm not mistaken, this camera is similar to the Viper, in that you can get your shot and do all the color correction (moreso than normal) and everything else later in post. Or, if you wish and have some extra time (and who does in the indie world?), you can get the majority of what people will see on screens during the shoot with minimal correction in post.

heath