Log in

View Full Version : Image bit planes show chequerboard patten in many macroblocks?


Yellow_
18th December 2011, 12:36
Looking for comments on the following two images extracted from native video files from two cameras and then a bit plane extracted from each.

http://www.yellowspace.webspace.virginmedia.com/C300-Tram_8-adj.png

http://www.yellowspace.webspace.virginmedia.com/T2i_MOV_8-adj.png

A Canon C300 '4k' $20,000 camera shooting XDCAM 422 in an MXF source files from here: http://ninofilm.net/blog/2011/12/09/canon-eos-c300-review-short-film/ which according to sources like dvinfo.net : http://www.dvinfo.net/article/acquisition/canon-eos/canon-usa-announces-cinema-eos-c300-and-eos-c300-pl-cameras.html

"The sensor has a resolution of 2,160 pixels tall by 3,840 pixels wide, which qualifies as native 4K.

...every four pixels (two green, one red, and one blue) are sampled for each final output pixel. In other words, color is assembled the same way as a traditional three-chip sensor block… two megapixels of red, two megapixels of blue and four megapixels of green (twice as much green as red or blue, since green carries the luminance info). Each primary color sampling off of the sensor is native 1920×1080, each color value alone is equal to the final output resolution. Canon claims that the processed signal has 1,000 lines of TV resolution, and the moire, diagonal line stair-stepping and other artifacts are greatly reduced in this chip compared to HD-DSLR cameras."

The interesting thing is the double luma and the description of the C300 giving a 'organic' pleasing grain, attractive digital noise :-), in the eye of the beyolder. :-)

The second image is the same bit plane level but extracted from a 'similar' scene in terms of exposure, colour and dof, but from a $500 Canon T2i / 550D DSLR h264AVC at 1920x1088.

FFmpegSource2 and ConvertToRGB PC levels were used to extract the image frames. Imagemagick was used for the bit plane extraction. Using Convert C300-Tram.png -fx '(int(u*255)&8)/255' C300-Tram_8.png and then Gimp to wack up the contrast and brightness by 127 to see the bit plane.

What interests me, :-) is the chequerboard pattern clearly seen on many of 'macroblocks?' in the bit planes from the C300 source, what could this be, it seems indiscriminate, is this bayer pattern? Ordered dithering? Any ideas?

It's not seen in the T2i / 550D source and assume this is a by product of the way the C300 sensor data is being put into a 1920x1080 frame.

Second query is probably going to be considered as pointless but with regard to 'cretinsealpes' Dither tools and LSB/MSB stacking and the possibility of taking the T2i/550D 1920x1088 h264 source up to 16bit by LSB/MSB, upsizing the luma to double res, maybe sharpening the luma somewhat, maybe sharpening by local contrast adjustment or USM instead, add a noise / grain layer via GrainFactory4 or Dither tools own grain functions or a dithering method that would allow further processing, all to bolster the luma plane, then resize back to 1920x1088 and add back chroma including upsampling of chroma to 4:2:2 and encode out to something like lossless 10bit h264 or similar for grading.

markanini
18th December 2011, 21:24
That marketing text seems more than a bit overblown to me (http://meanwhile.files.wordpress.com/2007/07/shreddies-ooh-02.jpg), alluding to 3CCD. And the "double luma" thing, it's the same type of Beyer filter avaliable in digicams since day one. The screenies looks like typical h264 macroblocking to me. Don't think it's the bayer filter as that would be well smoothed before reaching the encoder. The scenes look quite different not sure what one could get out of comparing them. Looks like a nice piece of kit.

Yellow_
18th December 2011, 23:02
Thanks for the response, :-) forgot to mention the XDCAM is mpeg2, which doesn't matter much in this regard.

But there are clearly chequerboard pattens to quite a number of 'macroblocks' in the C300 image that are not in the h264, may need to zoom to 200% to see them more clearly.

Of the 4 native C300 files and hundreds of h264, all the C300 files show the pattens in every frame, the h264 non of them.

Yeah the cameras getting good reviews, pre production units at the moment I think, release sometime early next year.

markanini
19th December 2011, 12:45
The "checkerboxes" might be unxpected if it's mpeg2. I'm no expert though. Those checkerboxes remind me of ordered dither, don't think I ever saw anything like that in MPEG2 video. Might be a codec implementation different from what I've seen? Or it is lacks an AA-filter and or the demosiac algorithm is'nt smooth. Could be a result of the decoder also.

Yellow_
20th December 2011, 10:44
Yes, my conclusion so far was that the chequerboards were a bug in the conversion, this is a pre production camera currently albeit immenent release in early new year and not the final codec implementation or choice possibly.

The other interesting thing was the natural grain / noise and whether that was a product of the bug or the product of the way the sensor data / noise is combined into the final 8bit 4:2:2 1920x1080 frame.

Here's an example of a luma key matte showing the grain:

http://blog.vincentlaforet.com/wp-content/uploads/Grain1.jpg