A fun day out with a 5D
Posted 13 December 2010 - 10:32 AM
Posted 13 December 2010 - 11:37 AM
How bad is the noise at full frame rate?
What Detail and ISO setting were you using on the 5D?
Posted 13 December 2010 - 02:26 PM
This particular one had pretty awful fixed-pattern noise, ....
Yes, it's bad enough around her mouth on the last image that I can even see it on this crappy computer. It's pretty clearly from the vertical decimation used on the 5D.
But everything else -- lighting, composition, makeup, wardrobe, hair, set dressing -- is all excellent.
Posted 13 December 2010 - 02:35 PM
Posted 18 December 2010 - 06:50 PM
Right, the decompress and UV upsample as done by Premiere Pro. Left, as converted to 10-bit uncompressed YUV by 5DtoRGB.
Almost all of the DSLR work that's currently being done is being converted by Quicktime, a program that uses Quicktime, or by a close equivalent, and will look like the picture at right. It is possible to do very much better.
Posted 18 December 2010 - 07:26 PM
FYI, Phil is using a development version of 5DtoRGB for Windows that does the image processing on the GPU, and the new code isn't fully tested. Results may improve/change later.
Posted 18 December 2010 - 09:07 PM
Just read up on fixed pattern noise (wikipedia ) - interesting stuff... Immediate thought was that you could fix it by shooting a test of blanks and comparing them to expected values, then subtract the differences later, etc... But in the usual nature of such things its all been done before and up to version 10 of the process, which is getting beyond me (and the ability for anyone writing a wikipedia page to describe without dolling out trade secrets).
What is the term to describe the artefacts on the lips ? I don't see how fixed pattern noise is the cause of that...
What I do note about those QT lips is that the effect on my macbook screen here disappears when I took down on the screen, and is gawd awful when I angle it to look up at it - probably saying more about my screen than anything right ?
Posted 18 December 2010 - 09:32 PM
This post assumes you know how subsampled YCbCr imaging works. If not, go read this.
What happens is that Quicktime (and others) do the upsample, well, quickly. This often looks worse than one might expect and I suspect other mistakes, with respect to interlacing and the quality of the initial downsample, are being made. This fits with the idea that the compression is being done on a DSLR with severe size, power consumption and cost limitations - reducing the size of an image (such as a Cr or Cb channel) and doing it well is hard work for computers, and this particular computer also has to do the h.264 compression, a very big job in itself. It looks very much as if the simplest possible downsample and upsample is being done, equivalent to "nearest neighbor" in Photoshop - just doubling pixels. This is massively primitive, and it is not actully technically that difficult for Thomas's software to do a much, much more careful job of upsampling the red and blue colour difference channels. It is slightly time consuming, which is why Quicktime doesn't do it, but then Quicktime is set up for realtime playback not post work. Quicktime also doesn't use GPU (that is, graphics card) processing power to do the job. Modern graphics rendering hardware, which is designed to render 3D scenes in realtime for computer games, are really vector processors (computer chips that run the same instructions over a large amount of data very rapidly) that are extremely well-suited to the sort of repetitive operations intrinsic to this sort of work, and versions of 5DtoRGB that do this are consequently very much faster than those that don't.
The results are plain to see.