I recently had the opportunity to attend a workshop with Charles Poynton, a renowned mathematician and video expert. The informal workshop took place in New York with just a few others, providing a great opportunity to listen and discuss hot button issue around High Dynamic Range (HDR).
Poynton prepared a thick handout of content and had this prepared as slides to review – but we may have looked at a half dozen during the whole day. The time was spent with Poynton verbally explaining the topics and fielding questions along the way. This can be an effective learning technique, but it ended up lacking the structure of a more formal workshop – which is valuable too.
He first presented his 0th Axiom of Digital Imaging (an axiom is self evident and can’t be proved the way a theorem can). This goes something like, “In the making of commercial images the only thing that matters is what happens at the approval process. Everything that happens downstream of this should not alter this image.” In other words, the “artistry” happens in mastering process and once approved, should be faithfully delivered to the end user.
Poynton hammered home some basics on light and light measurement. For example, he wanted us all to remember a few key light levels:
- 32,000 nits – luminance of white card illuminated by sunlight+skylight at noon on a clear day
- 32,000 nits plus 2 f-stops – headroom needed for specular highlights
- 320 nits – luminance of diffuse white rendered by a typical consumer TV
- 32 nits – luminance of typical diffuse white in cinema
- 3.2 nits – the light from a single candle at 1 foot distance
He next turned to focus on three significant effects. To illustrate the Hunt effect, Poynton suggested we think of the color of flowers in sunlight, then think of the color of those same flowers at twilight. The color of the flowers has not changed, but our perception of the colorfulness of them has – they don’t look as colorful in dim light. His point: “If you capture the flowers in daylight and show on a display with only 300 nits of brightness, the flowers will look like they are captured at twilight.” Since displays cannot show the full dynamic range of the natural world, we have to apply “artistic intent” to alter the image to produce what is desired. In other words, if you want the flowers rendered at 300 nits to have the visual colorfulness that they did in sunlight, you have to add more color. It is not accurate, but it conveys the artistic intent.
The Stevens effect is loosely comparable to the Hunt effect, but applies to visual contrast: “apparent contrast decreases with decreasing luminance.” Therefore, you also need to increase the contrast of the flowers example above to restore the perception of viewing them in sunlight.
The Bartleson & Breneman effect states that “grayscale tones surrounded by black appear visually different from the same tones surrounded by white.” This means the flowers described above would appear less colorful when surrounded by black than when surrounded by white – which to me seems counter intuitive. Since most content is mastered in a room with low level black surrounding light, this effect argues for adding more colorfulness to the image to compensate for this effect.
We ended up talking a lot about Perceptual Quantizer (PQ) and Hybrid Log Gamma (HLG) curves, why they are needed and how they work.
Poynton is well respected in the industry, and he has made a number of informative, but sometimes provocative or controversial statements. Below is a quick summary of some I remember.
- The color filters in an acquisition camera do not limit the color gamut of the capture content. Unfortunately, he never got around to explaining why this should be so as it is not obvious to me why that should be the case.
- Using Imperial units like footcandles [fc] and footlamberts [fL] impedes understanding of radiometry and photometry
- Confusion is rampant in ITU, SMPTE, EBU and MPEG standards regarding the reference white and peak white luminance levels. For SDR content, reference white is standardized by SMPTE (but not ITU) as code value 940 for a 10-bit word and 100 nits. Peak white is speced as code value 1019 according to ITU-R BT.1886, which equates to about 122 nits. There is inconsistent standardization on how to handle out of range luminance value and how to clip or not clip them – so there is inconsistent implementation of this.
- There is no definition of reference level for diffuse white for HDR, which is a big oversight. He recommends using 180 nits for diffuse white for UHD. Specular highlight should be up to 3X diffuse white for UHD (540 nits).
- He recommends reserving code values 64-940 in a 10-bit video for the reference data with 941-1019 for specular highlights
- Poynton uses the terminology OECF and EOCF instead of the more widely used OETF and EOTF
- Poynton would like to see the “brightness” adjustment we know eliminated from consumer TVs with the “contrast” knob relabeled “white level” and make sure it adjusts the diffuse white level, not the peak white level.
- HLG has been developed as a scene-referred OECF, but he thinks it should be explained as a display-referred transform the way PQ is.
- When referring to the EOCF (EOTF) use the notation PQ and HLG-1. When referring to the OECF (OETF) Poynton uses the notation PQ-1 and HLG to emphasize that one is scene-referred and the other is display-referred
- Poynton considers HDR10 to refer to 10-bit PQ or HLG-1 in a BT.2020 container (although there does not seem to be any document that states this explicitly)
- The MaxCLL (maximum content luminance) specification of ST-2086, which is part of the mastering metadata requirement, is based on Max [R’, G’, B’] and is not useful in determining maximum luminance, maximum display power or diffuse white luminance.
- The MaxFALL (maximum frame-average luminance level) has the same issue as it is based upon Max [R’, G’, B’]. It is not useful to determine the average or maximum luminance, average or maximum display power or diffuse white luminance.
It is hard to summarize an entire day of discussion in one article, but I think you get the idea. There remains a lot of work to be done on HDR/WCG to say the least.