Hi F360Fellows,
I have decided to take another small step toward a better understanding of how video (image) compression works, particularly in the encroaching era of 10-bit depth colour-capable hardware. Although colour gradation might be a second afterthought (not counting high-quality rendering, I would not dare to argue here, Mr TrippyLighting !), 10-bit, when combined with high resolution, should provide better line contrast, which is so important in CAD drawings.
As I indicated in my previous post, "Subjective assessments/judgments" are difficult to quantify, so I built a couple of utility tools making such evaluations more systematic.
In this chapter, I focused on assessing the "colour error" between the reference (a raw F360 output, unfortunately, 8-bit only) and the selected pixel colour formats for the output videos (in all cases, the libx265 codec was used). The PCFs (pixel chroma formats) were paired as 8-bit and 10-bit colour depths, respectively (yuv420p|yuv420p10le, rgb24|x2rgb10le), arranged (UL|UR, DL|DR).
Warning! The resulting visuals are at 6K resolution and "nominally" in 10-bit colour depth. How they will present on your display … is an open question I cannot answer.
I can only say that I can distinguish some phantom differences on a 10-bit depth UHD+ OLED display. If you see only black … DO NOT throw your monitor into the rubbish bin; instead, consider upgrading it.
View the attached files in the native environment after downloading them. Some are substantial in size, but I purposely refrained from compressing them. VLC video viewer, which is recommended, and GIMP can handle PNGs. Both software tools are free and in the public domain.
So, it is time for … a test of your monitor.
The PNG shows the difference between an image delivered by F360 and the corresponding frame from a video encoded with the same image. The PNGs are in 48-bit format (16-bit red, 16-bit green, 16-bit blue). Mind you, your system's internal graphics path will squeeze this 48-bit representation within its capabilities.
Perhaps histograms are the most visible differences between the tiles.
The histogram resolution is 6K[0]=6016, although other synthetic numeric data overlaid on the picture are scaled to a more understandable 8-bit range. Thus, the height of the leftmost histogram bin (clipped if necessary) represents the number of pixels with colour error of 1/65536 for each r|g|b channel. The population of errors-by-value drops sharply as we move to the right on the chart, with the range purposely limited to 1/8 of the full 255-colour domain, since larger errors are less likely.
Error distributions (histograms) vary between different codecs. One might assume that the steepest distribution and less lingering errors (to the right of the histogram) should yield "the sharpest" video. For me, the x2rgb10le is the winner, although yuv420 PCF – equivalent to a nominal 1.5 bytes per pixel should be the most suitable (from the selection) for 8-bit depth of 3 bytes of F360 colour output when accounting for the compression ratio. I have not tried YUV422 or YUV444, pixel compression formats (chromas), though. Likely, the preference for the PCF might differ between potential native 10-bit image input and the current, unfortunately, only 8-bit.
- threshold_26_420p_26_422p10le_26_rgb24_26_x2rgb10le_0029_AB_ovl_D.png
The above assessment held until I processed the second tile quadruple (yuv420p|yuv420p16le, yuv422p16le|yuv444p16le). I went full throttle there, choosing a 16-bit double-byte colour depth for the comparison (the first tile stays the same). The rough picture of the quad-charts is presented below, but only as a placeholder. The real 6K resource should be downloaded and analysed in the native environment of your display system.
- threshold_26_420p_26_420p16le_26_422p16le_26_444p16le_0029_AB_ovl_D.png
You might rightfully complain about your handicapped vision predicament and inability to see 2^16 colour graduations, but relax, you are not alone. In the not-too-distant future, you will be generously offered an obligatory opportunity to equip yourself with an appropriate vision implant, purchased on a low-interest 30-year credit plan. Then you will know what you cannot see clearly now. The future is bright! Just be patient!
In the meantime, to satisfy your 16-bit dream, the videos below (once again to be viewed in their native form), like pinholes, demonstrate how the world (in this case, video codecs' pixel colour errors) can be seen with ever-increasing visual bit depth.
- threshold_16_420p_26_422p10le_26_rgb24_26_x2rgb10le_x265.mp4
- threshold_26_420p_26_420p16le_26_422p16le_26_444p16le_x265.mp4
Short description of how even the tiniest pixel colour errors can be viewed.
First, the whole error map (array) is clipped to a range from 0 to the error threshold. The threshold value is indicated on the video frames and standardised to the <0, 255> range. Since the array values can range from 0 to 16 bits, the internal values are scaled accordingly, thus achieving a granularity of 1/65536th. In the second step, the result is normalised to an 8-bit visual range for 'the implant less' display. Thus, the first frames of a video represent the smallest pixel colour errors, which, as frames progress, are scaled to a wider range and become "fainter and fainter". The "pixel error sweep" is organised to emphasise both the small error range and the one approaching the 7-bit thresholds (it is not of the linear rate). Beyond 7 bits, the screen would be too dark to show any visible changes.
The dynamic histogram overlayed on each video frame depicts the error distribution. The most extreme bins are clipped.
The complementary PNG files show the absolute pixel colour error map at the full 8-bit thresholds, extending the 8-bit visual range as in the videos, with some occasional vision-impairment enhancements.
- threshold_26_420p_26_422p10le_26_rgb24_26_x2rgb10le_0029_AB_ovl_D.png
- threshold_26_420p_26_420p16le_26_422p16le_26_444p16le_0029_AB_ovl_D.png
So, although your (our) vision is still constrained to an 8-bit range, you (we) can peep through a hole and anticipate what's going on beyond… even without an implant. Yes, I know, it is a poor man's solution… but what can we do?
Here is a preview of how the separate 16-bit split, lower byte and upper byte of the future VisionVerse would be revealed before your enhanced eyes. Note, the first image in the post shows the convolution of both bytes, or of the two images below.
- threshold_26_420p_26_422p10le_26_rgb24_26_x2rgb10le_0029_AB_ovl_L.jpg
- threshold_26_420p_26_422p10le_26_rgb24_26_x2rgb10le_0029_AB_ovl_D.jpg
We agree, don't we? The future is bright… or dark, as in the current case. Your (our) current 8-bit perception will be only a tiny fraction (1/256th) of a much richer 16-bit world… almost insignificant… to be bragging about it… and the surrounding dark ages!
What about more serious conclusions and weaknesses of the analysis, lensed by the exercise? I already have some preliminary ones; they might even be revolutionary. As some know, revealing revelations in the dark ages face a risk. I will weigh it… until the next post… in the dark.
Attached files:
threshold_16_420p_26_422p10le_26_rgb24_26_x2rgb10le_x265.mp4 https://a360.co/4v5VNwC
threshold_26_420p_26_422p10le_26_rgb24_26_x2rgb10le_0029_AB_ovl_L.jpg https://a360.co/49sOqHk
threshold_26_420p_26_422p10le_26_rgb24_26_x2rgb10le_0029_AB_ovl_D.jpg https://a360.co/4u0ZHWK
threshold_26_420p_26_422p10le_26_rgb24_26_x2rgb10le_0029_AB_ovl_D.png https://a360.co/4u5oU25
threshold_26_420p_26_420p16le_26_422p16le_26_444p16le_x265.mp4 https://a360.co/4dAPM5q
threshold_26_420p_26_420p16le_26_422p16le_26_444p16le_0029_AB_ovl_L.jpg https://a360.co/4uF8E9o
threshold_26_420p_26_420p16le_26_422p16le_26_444p16le_0029_AB_ovl_D.jpg https://a360.co/4fJF6Tc
threshold_26_420p_26_420p16le_26_422p16le_26_444p16le_0029_AB_ovl_D.png https://a360.co/4uD1UsE
FlattenFace_26_rgb24.mp4 https://a360.co/4f6iVXa
FlattenFace_26_x2rgb10le.mp4 https://a360.co/4uuNRFf
FlattenFace_26_444p10le.mp4 https://a360.co/3RAl5oa
FlattenFace_26_444p.mp4 https://a360.co/3PDu3R3
FlattenFace_26_422p10le.mp4 https://a360.co/4tWPryw
FlattenFace_26_422p.mp4 https://a360.co/49onubG
FlattenFace_26_420p10le.mp4 https://a360.co/4af9uRR
FlattenFace_26_420p.mp4 https://a360.co/4e66fOW
FlattenFace_26_420p16le.mp4 https://a360.co/4dM8F3Q
FlattenFace_26_422p16le.mp4 https://a360.co/3Rwxj1i
FlattenFace_26_444p16le.mp4 https://a360.co/432wEXN
To be viewed on 4K media devices (monitors, UHD TVs, projectors...) of reasonable performance and preferably 10-bit colour depth capability. For the best experience, use a stand-alone media player (e.g., VLC) and set the native resolution to full screen. Download the files over a network where the cost of doing so is not a concern. The files are to be used for private, non-commercial purposes only.
Have the 16R16G16B⨂16α Night…..
Regards
MichaelT
MichaelT
