Video:Measuring Video Quality with VMAF – Why You Should Care

VMAF, from Netflix, has become a popular tool for evaluating video quality since its launch as an Open Source project in 2017. Coming out of research from the University of Southern California and The University of Texas at Austin, it’s seen as one of the leading ways to automate video assessment.

Netflix’s Christos Bampis gives us a brief overview of VMAF’s origins and its aims. VMAF came about because other metrics such as MS-SSIM and, in particular, PSNR aren’t close enough indicators of quality. Indeed, Christos shows that when it comes to animated content (i.e. anime and cartoons) subjective scores can be very high, but if we look at the PSNR score it can be the same as the PSNR of score another live-action video clip which humans rate a lot lower, subjectively. Moreover, in less extreme examples, Christos explains. PSNR is often 5% or so away from the actual subjective score in either direction.

To a simple approximation, VMAF is a method of bringing out the spatial and temporal information from a video frame in a way which emphasises the types of things humans are attuned to such as contrast masking. Christos shows an example of a picture where artefacts in the trees are much harder to see than similar artefacts on a colour gradient such as a sky or still water. These extraction methods take account of situations like this and are then fed into a trained model which matches the results of the model with the numbers that humans would have given it. The idea being that when trained on many examples, it can correctly predict a human’s score given a set of data extracted from a picture. Christos shows examples of how well VMAF out-performs PSNR in gauging video quality.

 

Challenges are in focus in the second half of the talk. What are the things which still need working on to improve VMAF? Christos zooms in on two: design dimensionality and noise. By design dimensionality, he means how can VMAF be extended to be more general, delivering a number which has a consistent meaning in different scenarios? As the VMAF model has been trained on AVC, how can we deal with different artefacts which are seen with different codecs? Do we need a new model for HDR content instead of SDR and how should viewing conditions, whether ambient light or resolution and size of the display device, be brought into the metric? The second challenge Christos highlights is noise as he reveals VMAF tends to give lower scores than it should to noisy sources. Codecs like AV1 have film-grain synthesis tools and these need to be evaluated, so behaving correctly in the presence of video noise is important.

The talk finishes with Christos outlining that VMAF’s applicability to the industry is only increasing with new codecs coming out such as LCEVC, VCC, AV1 and more – such diversity in the codec ecosystem wasn’t an obvious prediction in 2014 when the initial research work was started. Christos underlines the fact that VMAF is a continually evolving metric which is Open Source and open to contributions. The Q&A covers failure cases, super-resolution and how to interpret close-call results which are only 1% different.

Watch now!
Download the presentation
Speaker

Christos Bampis Christos Bampis
Senior Software Engineer,
Netflix

Video: Encoding Vs Compute Efficiency in Video Coding

Ioannis Katsavounidis from Facebook joins us to talk us through his work finding the best balance between computation and encoding. He explains how encoding has moved from real-time, hardware-based encoding in the late 80s and 1990s through to file encoding, chunk-based encoding and now shot-based encoding. Each of these stages has brought opportunities to speed up encoding, but there has always been a fundamental reason why encoding can’t simply be sped up by the advance of IT.

Moore’s law posits that every year, the number of transistors in chips doubles. Whilst this has continued to be true until recent years, transistors have always been a proxy for processing power. For many years now, the way to keep the computational ability of CPUs high has been not to increase clock-speed as it was twenty years ago, but to add cores to the chip. As each core acts as its own CPU, this gives the ability to execute code in parallel with a thread of code running separately on each core. Whilst 12-20 cores are typical for servers, there are CPUs which deliver up to 128 cores.

Ioannis explains why DCT-based codecs are resistant to multi-thread encoding by showing how some of the encoding decisions are based on the previously decoded video frame so the encoder needs to decode the video before it has the information it needs to make the next encode decisions. An example of this motion estimation where you need to understand what a macroblock looks like in order to detail if and how it can be moved to form part of the macroblock currently being encoded.

It turns out that some of the information you need to calculate can be found from the original video. Whilst this doesn’t provide full parallelisation, it does help in freeing some of the computation to be done in parallel thus reducing the length of time spent on the linear encoding stage. As the design of the codec itself is limited in its ability to be parallelised, the best way to speed up encoding has been to split up the original video and encode these, now separate, sections independently.

Speeding up video encoding has therefore focused on splitting up the video into different sections and encoding those in parallel rather than trying to parallelise the encoding itself due. Encoding each frame separately is one way to do this, but sacrifices encoding efficiency. Splitting each frame up into sections (tiles or slices) is another way, though this also sacrifices either quality or bitrate. The most successful encoding parallelisation has been chunked encoding. As streaming applications use chunks, typically around 2 seconds nowadays, there’s no reason not to just cut your video up into small sections and encode those separately; the whole of this video focuses on non-live video.

If there’s a shot change in the middle of your chunk, this is likely to look very bad since the motion estimation will fail to produce good results and there may not be enough bitrate budget to compensate. Therefore it’s best to drop in an IDR frame at the shot change or to actually change your video chunks to match shot changes. Simply encoding these chunks in parallel would speed up the encoding, however, it misses an opportunity to optimise quality vs bitrate.

Ioannis explains an experiment to determine the best operating point for chunks. He does that by reminding us that all encoders have certain ‘speed’ settings which control how much computation, and therefore time, is required for each encode. The ‘very fast’ setting in x264 will encode at the highest speed possible, but the quality will be worse or a certain bitrate compared to the ‘very slow’ setting. Ioannis’s experiment encoded each chunk at every speed setting for a variety of resolutions and bitrates. Each encode was then analysed for quality using PSNR, MS-SSIM and VMAF.

From Ioannis’ work, we can see how the bitrate setting affects both the encode time and the quality and we can observe that the slower speeds tend to have minimal quality advantages for the significant extra time involved in the encoding. Each curve has a steep part and a shallow section with the transition between known as the ‘convex hull’. Choosing a setting on the convex hull portion of the line is the optimal balance between quality and encoding time and is where, says Ioannis, most people should aim to operate.

The talk finishes with a summary of the conclusions which can be drawn from this work looking at the use of convex-hull which we’ve just discussed, the best type of parallel processing, whether oversubscription of CPU cores is helpful or not and an interesting observation that it’s often the metrics which put a significant burden on encoding rather than the video encoding itself, particularly for lower resolutions.

Watch now!
Speakers

Ioannis Katsavounidis Ioannis Katsavounidis
Research Scientist,
Facebook

Video: Scaling up Anime with Machine Learning and Smart Real Time Algorithms

Too long has video been dominated by natural scenes and compression has been about optimising for skin tones. Recently we have seen technologies taking care of displaying other types of video correctly like computer displays such as computer games, as seen in VVC and also animation optimisation for upscalers as we explore in this talk.

Anime, a Japanese genre of animation, is not very different from an objective point of video from most video cartoons; the drawing style is black lines on relatively simple, solid areas of colour. Anime itself is a clearly distinct genre whose fans are much more sensitive to quality, but for codecs and scalers, 2D animation, in general, is a style that easily shows artefacts.

Up- and down-scaling is the process of making an image of say 1080 pixels high and 1920 wide larger, for instance 2160×3840 or smaller, say to SD resolution. Achieving this without jagged edges or blurriness is difficult and conventional maths can do a decent job, but often leaves something to be desired. Christopher Kennedy from Crunchyroll explains the testing he’s done looking at a super resolution upscaling technique which uses machine learning to improve the quality of upscaled anime video.

Waifu2x is an opensource algorithm which uses Convolutional Neural Networks (CNNs) to scale images and remove artefacts. To start with, Christopher explains the background of traditional algorithmic upscaling discussing the fact that better-looking algorithms take longer so TVs often choose the fastest leading them to look pretty bad if fed SD video. Better for the streaming provider to spend the time doing an upconversion to 4K so allow the viewer a better final quality on their set.

Machine Learning needs a training set and one thing which has contributed to waifu2x’s success in Anime is that it has been trained only on examples of anime leaving it well practised in improving this type of image. Christopher presents the results of his tests comparing standard bilinear and bicubic scaling with waifu2x showing the VMAF, PSNR and SSIM scores.

Finishing off the video, Christopher talks about the time this waifu2x takes to run, the cost of running it in the cloud and he shares some of the command lines he used.

Reference links:

Watch now!
Speaker

Christopher Kennedy Christopher Kennedy
Staff Video Engineer,
Crunchyroll

Video: Banding Impairment Detection

It’s one of the most common visual artefacts affecting both video and images. The scourge of the beautiful sunset and the enemy of natural skin tones, banding is very noticeable as it’s not seen in nature. Banding happens when there is not enough bit depth to allow for a smooth gradient of colour or brightness which leads to strips of one shade and an abrupt change to a strip of the next, clearly different, shade.

In this Video Tech talk, SSIMWAVE’s Dr. Hojat Yeganeh explains what can be done to reduce or eliminate banding. He starts by explaining how banding is created during compression, where the quantiser has reduced the accuracy of otherwise unique pixels to very similar numbers leaving them looking the same.

Dr. Hojat explains why we see these edges so clearly. By both looking at how contrast is defined but also by referencing Dolby’s famous graph showing contrast steps against luminance where they plotted 10-bit HDR against 12-bit HDR and show that the 12-bit PQ image is always below the ‘Barten limit’ which is the threshold beyond which no contrast steps are visible. It shows that a 10-bit HDR image is always susceptible to showing quantised, i.e. banded, steps.

Why do we deliver 10-bit HDR video if it can still show banding? This is because in real footage, camera noise and film grain serve to break up the bands. Dr. Hojat explains that this random noise amounts to ‘dithering’. Well known in both audio and video, when you add random noise which changes over time, humans stop being able to see the bands. TV manufacturers also apply dithering to the picture before showing which can further break up banding, at the cost of more noise on the image.

How can you automatically detect banding? We hear that typical metrics like VMAF and SSIM aren’t usefully sensitive to banding. SSIMWAVE’s SSIMPLUS metric, on the other hand, has been created to also be able to create a banding detection map which helps with the automatic identification of banding.

The video finishes with questions including when banding is part of artistic intention, types of metrics not identifiable by typical metrics, consumer display limitations among others.

Watch now!
Speakers

Dr. Hojat Yeganeh Dr. Hojat Yeganeh
Senior Member Technical Staff,
SSIMWAVE Inc.