Your browser is not supported. Please update it.

12 July 2018

Akamai warns against lowering bit rate to avoid buffering

It is common practice for streaming services to reduce bit rate at points and times of congestion to avoid buffering, but this can defeat the object of improving the overall viewing experience, according to CDN provider Akamai. Service providers tend to regard lowering the bit rate temporarily to sacrifice some picture resolution as a lesser evil than buffering, which turns viewers off quickly and is a major cause of churn in OTT.

But excessive use of bit rate reduction means that viewers expecting say HD quality are often having to make do with SD or even worse for a significant proportion of the running time. This too can provoke churn even if it is more of a slow burn, especially now that expectations of OTT quality have converged with broadcast, whereas a few years ago consumers were more tolerant of glitches or deteriorations in resolution. With HD now defacto SD has become offensive to the eye.

This trend towards ever higher picture expectations has implications for measurements of perceptual quality which has become a growing issue both for providers of services and the underlying infrastructure. Akamai has become aware of the importance of delivering consistent perceptual quality and also of the fact that the bit rate required to deliver it varies significantly between content types. Accordingly it has teamed up with Eurofins, a Brussels based group of analytical laboratories spanning multiple disciplines, to assess this relationship between bit rate, content type, screen format and perceptual quality for streaming services. That is four variables with the main challenge being how to measure perceptual quality against this background of inflationary expectations.

There is always direct human assessment to yield Mean Opinion Scores (MOS) but that clearly does not scale up to services delivering large amounts of content of varying types. So the baseline technical alternative widely used until quite recently has been PSNR (Peak Signal to Noise Ratio), which is a raw measure of the visual information getting through by taking account of the amount of noise corrupting the signal. Although widely derided, often by those promoting more sophisticated techniques, PSNR provides a reasonably accurate baseline measure of visual quality. However it has looked increasingly unfit for purpose as quality has improved with the advent of Ultra HD. As a result Netflix developed its open source Video Multimethod Assessment Fusion (VMAF), which was a heavy data collection and crunching exercise to match human perception with underlying transmission characteristics, involving non-experts ranking samples of content from 1 to 5.

But Netflix is only interested in on-demand delivery via TCP which means that dropped packets and bit errors are not a factor, just compression artefacts and scaling artifacts when video is downsampled before compression to cater for lower bit rate networks and later upsampled on the viewer’s device. In both those cases information is lost on the way and the question is whether this has a noticeable impact on quality.

But since VMAF was optimized only for TCP it might not work so well for live transmission where there is not time to resend packets and so the connectionless UDP is often used. This did not rule out VMAF though and so tests have been conducted in various labs with a few alternatives, notably SSIMplus from SSIMwave, a Canadian specialist in video quality assessment. This was found to be 49% more accurate and 10 times more computationally efficient than VMAF in a 2017 study by the University of Waterloo in Ontario. This method employed the regressive methods under the banner of machine learning and some other mathematical difference techniques to hone the measurements to perceptions reported by humans on samples, ranking content on a scale 0 to 100.

The Akamai-Eurofins study applied SSIMplus to various types of content on different smartphones, set tops, desktop PCs, tablets and smart TVs. For this purpose any score of 80 or above was deemed at least equivalent to HD, irrespective of the viewing device.

Three different content types were tested, talking heads, drama and high action content including sports. These were matched against the various combinations of player, content and bitrate, tested under both normal and challenging network conditions where available bandwidth varied constantly.
The results showed greater variation in the bitrate required to deliver a consistent HD image than expected and also significant differences between the capabilities of the devices to mitigate the issue themselves. The ratio between bit rate required to maintain HD quality on smart TVs for fast action content and that for talking heads was at least three to one. Smart TVs required 6 Mbps to 8 Mbps for high action content, 2 Mbps to 4 Mbps for drama and less than 2 Mbps for talking heads.

Desktop PCs required 5 Mbps to maintain HD given fast moving content and smartphones just under 3 Mbps, with again significant differences between the genres. The set top though outperformed the smart TV, especially under challenging network conditions, while the two smartphones tested exhibited significant perceptual quality variation at a given bit rate. This highlights how devices of a given screen size can deliver different perceptual qualities at identical bit rates and network conditions.

One surprising aspect of the tests as reported is lack of attention to content aware encoding (CAE), which as we have explained before has great potential to reduce bit rate for a given perceptual quality but dependent on content type. That is why, as we highlighted last week, the Ultra HD Forum has added CAE to its list of core technologies for its Phase B specifications for UHD services, recognizing its essential role in enabling for example High Frame Rate at 100 fps or more without busting bandwidth budgets.

A key point is that even for high action content CAE can reduce bit rates by around 23% on top of the compression achieved by codecs such as HEVC without reducing perceptual quality. It does this by exploiting the fact that even within the high action content there are some sub-sequences coming down almost to the frame level where not so much is happening, so they can be compressed further while sustaining perceptual quality.

In the case of Akamai’s study CAE would on that basis have reduced the bit rate required for high action HD content on smart TVs to well under 5 Mbps. CAE will be an essential part of the mix for delivering HD and especially UHD streaming services.