Wednesday, January 21, 2009

Correction: Canon 5D is fine, tools are wrong.

In my last post I discovered that using different decoders produce different results on Canon 5D footage, some more favorable than others, but I was wrong in my conclusion that something was being misinterpreted in the Canon 5D bit-stream. The Canon 5D video bit stream is fine, in fact is a little better that, as it making the best use of its heavily compressed 8-bit MOV (of course a less compressed 10-bit would have been nice.)

I concluded this by directly examining the luma outputs of two popular AVC/H.264 decoders: MainConcept and CoreAVC (both tested by snooping the YUV output via DirectShow.) Those users with CoreAVC were finding the output data easy to color correct, seeing the full dynamic range, whereas many MainConcept users (standard within Adobe and Sony NLEs) much less so. See these two histograms below explain this issue.

First CoreAVC. The luma data ranged between the broadcast standards of 16 to 235, not a single value existed out of this range (which is odd for most YUV source.) However, this is why CoreAVC users are having no big issues with dynamic range, even when using 8-bit RGB decodes, which will map the 16-235 luma to 0-255 in RGB -- they get everything. However, the spikes in the histogram are a tell-tale sign that something else is wrong with this picture.

Compare this to the Mainconcept decoder's output (same output as QT and other common AVC decoders), the completely smooth histogram over the full range from 0 to 255, shows that CoreAVC is in fact post compressing the data into the broadcast 16-235 range, and that is not how the Canon 5D compressed the image. This is range compression is reducing the tonal range, and the remapping of 8-bit values into a smaller 8-bit range can introduce banding. Think of a sky gradient with values 10,11,12,13,14,15 being compressed to 10,11,12,(12),13,14 -- the flat spot can add to visible contouring for which 8-bit signal are already prone.

So if the default decoders are correct, but dynamic range is truncated in the NLE, what is the solution? Stop using 8-bit RGB tools and codecs helps. The problem only arises when remapping the 8-bit YUV data into 8-bit limited computer graphics RGB space (Studio RGB would be fine *.) If your application and intermediate codec can decode to YUV, you are in fine shape, the problem is none of the common NLEs are doing this with native AVC/H.264 decoding. Intermediate codecs like ProRES and CineForm, support full range YUV (float) precising within FCP and Premiere Pro, so those are good routes to go. So you need to convert from the Canon MOV outside of NLE, using tools that do all processing in YUV (image remains unclipped.) Stu Maschwitz has posted that Apple Color will support the full range, and I can happily report that all CineForm products are handling this correctly (even for ProRes output if you are on a Mac with NEO HD/4K.) But even a ProRes file will clip if 8-bit RGB tools are used.

CineForm a cool solution for this using NEO HD/4K for the Mac and Prospect HD/4K on the PC, even if you have to use an 8-bit RGB clipping tools like Compressor, QT Player or VirtualDub, etc. The full range 0-255 YUV is preserve in the CineForm MOV or AVI, but using Active Metadata you can choose to decode the smaller range when needed. As this is all happening with high precision math in the decoder, we can safely mapping the full range YUV to full ragne RGB, without truncation or introducing banding. Normally we using Active Metadata for attaching look to the decode stream, but it can also use it to reformat the data to fix the needs of your tools.

The first RGB parade is the untouched CineForm encode from the MainConcept AVC decoder output, it appears that the shadows and highlight have been crushed. While I could place a 32-bit Levels filter to correct it on the timeline when work with the CineForm encoded file, correcting it within Active Metadata will allow this clip to also be corrected in 8-bit tools.
The second RGB parade is corrected through Active Metadata alone. I simply bumped up the black level and reduced the gain inside the Active Metadata control panel (Mac or PC) to reveal all the missing data. Active Metadata uses the clip's GUID (Global Unique IDentifier) to set new decoding rules, for any application use that clip with the computer system. Expect there to be more problems solved using Active Metadata with our next major software release. So much cool stuff coming -- need more software engineers.

* - At the time of the post Apple released a patch to QuickTime (7.6) that addresses some of the this. Now under QT 7.6 the 5D YUV data is decoded using Studio RGB, which places luma 0 (not 16) at RGB 0,0,0, allowing RGB to have the same gamut of the YUV source. I imagine the 5D is flags the data to do this and various AVC decoders are either ignoring or misusing the flag.

----

01/22/2009 It turns out the CoreAVC control has an auto mode, trying to guess the correct decode behavior (16-235 vs 0-255). Turn this off, make it behave just like all other. So all these decoders work if you manage the outputs correctly.

Monday, January 19, 2009

Full dynamic range video from Canon 5D MkII.

I been seeing many posts on how crushed the black levels and highlights are with the Canon 5D video mode. Stu Maschwitz has a great post on how to tweak the camera to improve the situation, but also points out the default presentation FCP is also wrong. It turns out that this presentation fault is in all popular editing packages on Mac and PC. They could possibly share the same AVC decoder, but I have a different theory below.

Using a 5D MOV video originally posted DPReview.com, here you can see the clipping that occurs within tools like Vegas, Premiere Pro and FCP when directly opening the MOV files. This zoomed in section shows much of the tree details are crushed into the blacks, when compared to the same sequence converted to an AVI using HDLink with CineForm NEO HD.

After some image enhancement, the black truncation becomes more apparent. These images where manipulated with a 32-bit levels filter (it is not truncating the data) so we see how much data is being lost, a failure to see "the trees from the forest" type of problem.

One common reason for all this error, is the camera H.264 I-frame compression is using 4:2:0 YUV, where display black is 16 and white is 235 out of a 0-255 luma range. If the NLE took this YUV data directly, everything should be fine, but instead is likely going through series YUV to RGB transforms that are truncating the data. Most NLEs extracting RGB from YUV sources map the 16-235 YUV to 0 to 255 RGB, truncating data in the supers, yet it turns out the Canon 5D doesn't even use much in super range (blacks between 0-15 and whites up to 255) as this non-truncated luma histogram shows: It seems that the 5D is limiting the output to broadcast standards (yet another thing the 5D does wrong for video heading for post.) Whereas most video cameras do use the extra range, they assume the broadcast limiting will be handled in post (as they should.) So the NLE's mapping of the 16-235 to 0-255 RGB is actually not the issue, somehow the H.264/AVC decoder used is actually mapping a range like 30-220 only into the NLE space. I have not seem such a big error before, which makes me think there an incorrect flag in the 5D bit-stream that is causing the use of the wrong output math.

Luckly within CineForm NEO HD we use direct YUV to YUV conversion wherever possible, this eliminates many trunction errors, and preserves highlight for mulitple camera sources. In current version of NEO/Prospect HD (v3.4) we don't ship an AVC decoder, so NEO HD searches out for registered decoders already installed, resulting in some users reporting this error and others not so. So the failure is AVC decoder dependent, although this may not be a bug in the decoders, as they all handle regular AVCHD sources without issue (again leading me to believe there is something wrong/different in the 5D bit-stream.) The one AVC decoder that produces the full range of YUV range is CoreAVC. We have been recommending it to AVCHD users as it is both fast and inexpensive (only $15), but it seems it might have extra value for those trying to get the most out of their Canon 5Ds.

In our next major release of NEO HD (and up) we intend to direct support Canon 5D, yet in the meantime for those using CineForm products I recommend you get CoreAVC, the latest version of NEO or Prospect which now favors CoreAVC over any other AVC decoder it might find. This version is currently in public beta and will likely become official in a day or so.

------
1/21/09
Update 1: The new software is now released and available for upgrade or trial directly from cineform.com.

Update 2: I have discover more on AVC Decoders issues for the Canon 5D in the next post.

Tuesday, January 13, 2009

CineForm's Winning Streak

Not only did Slumdog Millionaire clean up at the Golden Globes this year, Slumdog being a mostly CineForm RAW acquired feature shot with the Silicon Imaging SI-2K Mini, one of CineForm's engineers just won a Audi RS4 (Audi Record the Rush Challenge) from a CineForm posted video he self produced also shot with an SI-2K and a Sony V1U.  Looking to be a good 2009 for those at CineForm.