Disclaimers: Credit for images is given where possible, apologies for any omissions The optical demonstrations slides may not work on the target monitor / projector The HDR images have been tonemapped to show the effect on an SDR display The original work can be found here: https://mediasolutions.ericsson.com/cms/wpcontent/uploads/2017/10/ibc-2017-0316.pdf Why CbCr? Are current broadcast containers suitable for HDR WCG content? Olie Baumann Alex Okell Jacob Ström 1
Cameras, TVs and Eyes vision doctor.com askabiologist.asu.edu But Broadcast Is Different For broadcast we use a different container: Y CbCr Why? Backward compatibility with black and white TV Compression vision doctor.com askabiologist.asu.edu 2
Containers and Colour Spaces Container: A system for representing 3 tuple vectors RGB and Y CbCr are containers The same information can be stored in any container Colour space: Defines how 3 tuples represent physical colours Defined in a coordinate system CIE 1931 Covers more than the visible spectrum RGB to YCbCr ~. bitjazz.com 3
Compression Efficiency is improved in two ways: Natural scenes have significant correlation between R,G and B Y CbCr encapsulates the correlation in Y The human visual system has a greater acuity in luminance than chrominance We can transmit band limited Cb and Cr AKA chroma down sampling, 4:2:2, 4:2:0 4
Compression Efficiency is improved in two ways: Natural scenes have significant correlation between R,G and B Y CbCr encapsulates the correlation in Y The human visual system has a greater acuity in luminance than chrominance We can transmit band limited Cb and Cr AKA chroma down sampling, 4:2:2, 4:2:0 Example Glass-to-Glass Workflow 5
HD vs UHD (HDR) Colour Volumes Courtesy of Dolby HDR, WCG in Y CbCr It s a bigger volume (in xyy) But we want to use existing broadcast infrastructure (Y CbCr) So, what to do? Squeeze the dynamic range Increase the bit depth to 10 bit 6
Non-linear Transfer Functions Luminance Sensitivity 50 254 54 250 7
Non-linear Transfer Functions More quantisation levels for lower luminance levels Squeezes the colour volume imperceptibly At what cost? Error / noise amplification Increased correlation, Y no longer represents Y Error Amplification 8
What Does it Mean? Significant distortion in plain(ish) areas Chroma resampling 4:4:4 > 4:2:0 > 4:4:4 No other compression First reported in MPEG Original 4:2:0 Resampled Chroma Resampling Worked Example Two neighboring pixels: 1000, 0, 100 x 0.64, 0.26, 0.027 1000, 4, 100 x 0.63, 0.26, 0.027,, 0,10000 9
Chroma Resampling Worked Example Two neighboring pixels: 1000, 0, 100 263, 646, 831 1000, 4, 100 401, 571, 735 Chroma Resampling Worked Example Two neighboring pixels: 263, 1000, 646, 0, 831 100 401, 1000, 571, 4, 735 100 10
Chroma Resampling Worked Example Two neighboring pixels: 263, 646, 831 Simple average of chroma 263, 608.5, 783 401, 571, 735 401, 608.5, 783 Chroma Resampling Worked Example Two neighboring pixels: 1000, 0, 100 Reconstruction from : 484, 0.03, 45 1000, 4, 100 2061, 2.2, 216 11
Chroma Resampling Worked Example Two neighboring pixels: 0.64, 0.26, 0.027 Reconstruction from : 0.64, 0.26, 0.013 0.63, 0.26, 0.027 0.63, 0.26, 0.056 Errors in chroma affect the output luminance! But not everywhere! Characterising the Error If not everywhere, then where? Which colors are most affected? Not which colors but which pairs of colors 6 dimensional error Take a statistical approach Original 4:4:4 4:2:0 Resampled 12
Simulated Textures 16 by 16 pixel swatch Base color plus noise Base color taken on a grid defined in Process according to the glass to glass workflow PQ (SMPTE ST. 2084) OETF / EOTF BT. 2020 color space Calculate the error (SSD) in luminance, Results 13
Results Changes with Luminance 14
Characterising the Error The hues most affected change as a function of the luminance Errors are greatest at the edge of the colour space Most of the errors are outside the BT. 709 range Which implies that these artefacts are likely to become more common as cameras capture more of the BT. 2020 colour volume What About HLG? 15
What About HLG? Results: HLG peak 1000 cd/m 2 16
Results: HLG vs PQ Results: HLG vs PQ 17
Results: HLG vs PQ What can we do About it? Avoid the edge of the colour gamut Move to a different container Constant luminance / constant intensity Specific pre processing Ericsson s chroma correction in the downsampling for example 18
Conclusions The use of more non linear transfer functions further decorrelates luma ( ) from luminance ( ) Errors in and spill over into the luminance These errors have greatest effect at the edges of the color gamut This impacts how we treat chroma in all processing performed on video It will become more of an issue as cameras capture more of the full BT. 2020 color gamut 19