Welcome to the fourth column on the ACM SIGMM Records from the Video Quality Experts Group (VQEG).
During the last VQEG plenary meeting (14-18 Dec. 2020) various interesting discussions arose regarding new topics not addressed up to then by VQEG groups, which led to launching three new sub-projects and a new project related to: 1) clarifying the computation of spatial and temporal information (SI and TI), 2) including video quality metrics as metadata in compressed bitstreams, 3) Quality of Experience (QoE) metrics for live video streaming applications, and 4) providing guidelines on implementing objective video quality metrics to the video compression community.
The following sections provide more details about these new activities and try to encourage interested readers to follow and get involved in any of them by subscribing to the corresponding reflectors.
SI and TI Clarification
The VQEG No-Reference Metrics (NORM) group has recently focused on the topic of spatio-temporal complexity, revisiting the Spatial Information and Temporal Information (SI/TI) indicators, which are described in ITU-T Rec. P.910 . They were originally developed for the T1A1 dataset in 1994 . The metrics have found good use over the last 25 years – mostly employed for checking the complexity of video sources in datasets. However, SI/TI definitions contain ambiguities, so the goal of this sub-project is to provide revised definitions eliminating implementation inconsistencies.
Three main topics are discussed by VQEG in a series of online meetings:
- Comparison of existing publicly available implementations for SI/TI: a comparison was made between several public open-source implementations for SI/TI, based on initial feedback from members of Facebook. Bugs and inconsistencies were identified with the handling of video frame borders, treatment of limited vs. full range content, as well as the reporting of TI values for the first frame. Also, the lack of standardized test vectors was brought up as an issue. As a consequence, a new reference library was developed in Python by members of TU Ilmenau, incorporating all bug fixes that were previously identified, and introducing a new test suite, to which the public is invited to contribute material. VQEG is now actively looking for specific test sequences that will be useful for both validating existing SI/TI implementations, but also extending the scope of the metrics, which is related to the next issue described below.
- Study on how to apply SI/TI on different content formats: the description of SI/TI was found to be not suitable for extended applications such as video with a higher bit depth (> 8 Bit), HDR content, or spherical/3D video. Also, the question was raised on how to deal with the presence of scene changes in content. The community concluded that for content with higher bit depth, SI/TI functions should be calculated as specified, but that the output values could be mapped back to the original 8-Bit range to simplify comparisons. As for HDR, no conclusion was reached, given the inherent complexity of the subject. It was also preliminarily concluded that the treatment of scene changes should not be part of an SI/TI recommendation, instead focusing on calculating SI/TI for short sequences without scene changes, since the way scene changes would be dealt with may depend on the final application of the metrics.
- Discussion on other relevant uses of SI/TI: it has been widely used for checking video datasets in terms of diversity and classifying content. Also, SI/TI have been used in some no-reference metrics as content features. The question was raised whether SI/TI could be used for predicting how well content could be encoded. The group noted that different encoders would deal with sources differently, e.g. related to noise in the video. It was stated that it would be nice to be able to find a metric that was purely related to content and not affected by encoding or representation.
As a first step, this revision of the topic of SI/TI has resulted in a harmonized implementation and in the identification of future application areas. Discussions on these topics will continue in the next months through audio-calls that are open to interested readers.
Video Quality Metadata Standard
Almost all modern transcoding pipelines use full-reference video quality metrics to decide on the most appropriate encoding settings. The computation of these quality metrics is demanding in terms of time and computational resources. In addition, estimation errors propagate and accumulate when quality metrics are recomputed several times along the transcoding pipeline. Thus, retaining the results of these metrics with the video can alleviate these constraints, requiring very little space and providing a “greener” way of estimating video quality. With this goal, the new sub-project has started working towards the definition of a standard format to include video quality metrics metadata both at video bitstream level and system layer .
In this sense, the experts involved in the new sub-project are working on the following items:
- Identification of existing proposals and working groups within other standardisation bodies and organisations that address similar topics and propose amendments including new requirements. For example, MPEG has already worked on the adding of video quality metrics (e.g., PSNR, SSIM, MS-SSIM, VQM, PEVQ, MOS, FISG) metadata at system level (e.g, in MPEG2 streams , HTTP , etc.).
- Identification of quality metrics to be considered in the standard. In principle, validated and standardized metrics are of interest, although other metrics can be also considered after a validation process on a standard set of subjective data (e.g., using existing datasets). New metrics to those used in previous approaches are of special interest. (e.g., VMAF , FB-MOS ).
- Consideration of the computation of multiple generations of full-reference metrics at different steps of the transcoding chain, of the use of metrics at different resolutions, different spatio-temporal aggregation methods, etc.
- Definition of a standard video quality metadata payload, including relevant fields such as metric name (e.g., “SSIM”), version (e.g., “v0.6.1”), raw score (e.g., “0.9256”), mapped-to-MOS score (e.g., “3.89”), scaling method (e.g., “Lanczos-5”), temporal reference (e.g., “0-3” frames), aggregation method (e.g., “arithmetic mean”), etc .
More details and information on how to join this activity can be found in the NORM webpage.
QoE metrics for live video streaming applications
The success of a live multimedia streaming session is defined by the experience of a participating audience. Both the content communicated by the media and the quality at which it is delivered matter – for the same content, the quality delivered to the viewer is a differentiating factor. Live media streaming systems undertake a lot of investment and operate under very tight service availability and latency constraints to support multimedia sessions for their audience. Both to measure the return on investment and to make sound investment decisions, it is paramount that we be able to measure the media quality offered by these systems. In this sense, given the large scale and complexity of media streaming systems, objective metrics are needed to measure QoE.
Therefore, the following topics have been identified and are studied :
- Creation of a high quality dataset, including media clips and subjective scores, which will be used to tune, train and develop objective QoE metrics. This dataset should represent the conditions that take place in typical live media streaming situations, therefore conditions and impairments comprising audio and video tracks (independently and jointly) will be considered. In addition, this datasets should cover a diverse set of content categories, including premium contentes (e.g., sports, movies, concerts, etc.) and user generated content (e.g., music, gaming, real life content, etc.).
- Development of QoE objective metrics, especially focusing on no-reference or near-no-reference metrics, given the lack of access to the original video at various points in the live media streaming chain. Different types of models will be considered including signal-based (operate on the decoded signal), metadata-based (operate on available metadata, e.g. codecs, resolution, framerate, bitrate, etc.), bitstream-based (operate on the parsed bitstream), and hybrid models (combining signal and metadata) . Also, machine-learning based models will be explored.
Certain challenges are envisioned to be faced when dealing with these two topics, such as separating “content” from “quality” (taking int account that content plays a big role on engagement and acceptability), spectrum expectations, role of network impairments and the collection of enough data to develop robust models . Readers interested in joining this effort are encouraged to visit AVHD webpage for more details.
Implementer’s Guide to Video Quality Metrics
In the last meeting, a new dedicated group on Implementer’s Guide to Video Quality Metrics (IGVQM) was set up to work on introducing and provide guidelines on implementing objective video quality metrics to the video compression community.
During the development of new video coding standards, peak-signal-to-noise-ratio (PSNR) has been traditionally used as the main objective metric to determine which new coding tools to be adopted. It has been furthermore used to establish the bitrate savings that a new coding standard offers over its predecessor through the employment of the so-called “BD-rate” metric  that still relies on PSNR for measuring quality.
Although this choice was fully justified for the first image/video coding standards – JPEG (1992), MPEG1 (1994), MPEG2 (1996), JPEG2000 and even H.264/AVC (2004) – since there was simply no other alternative at that time, its continuing use for the development of H.265/HEVC (2013), VP9 (2013), AV1 (2018) and most recently EVC and VVC (2020) is questionable, given the rapid and continuous evolution of more perceptual image/video objective quality metrics, such as SSIM (2004) , MS-SSIM (2004) , and VMAF (2015) .
This project attempts to offer some guidance to the video coding community, including standards setting organisations, on how to better utilise existing objective video quality metrics to better capture the improvements offered by video coding tools. For this, the following goals have been envisioned:
- Address video compression and scaling impairments only.
- Explore and use “state-of-the-art” full-reference (pixel) objective metrics, examine applicability of no-reference objective metrics, and obtain reference implementations of them.
- Offer temporal aggregation methods of image quality metrics into video quality metrics.
- Present statistical analysis of existing subjective datasets, constraining them to compression and scaling artifacts.
- Highlight differences among objective metrics and use-cases. For example, in case of very small differences, which metric is more sensitive? Which quality range is better served by what metric?
- Offer standard logistic mappings of objective metrics to a normalised linear scale.
 ITU-T Rec. P.910. Subjective video quality assessment methods for multimedia applications, 2008.
 M. H. Pinson and A. Webster, “T1A1 Validation Test Database,” VQEG eLetter, vol. 1, no. 2, 2015.
 I. Katsavounidis, “Video quality metadata in compressed bitstreams”, Presentation in VQEG Meeting, Dec. 2020.
 I. Katsavounidis et al. “A case for embedding video quality metrics as metadata in compressed bitstreams”, working document, 2019.
 ISO/IEC 13818-1:2015/AMD 6:2016 Carriage of Quality Metadata in MPEG2 Streams.
 ISO/IEC 23009 Dynamic Adaptive Streaming over HTTP (DASH).
 ISO/IEC 23001-10, MPEG Systems Technologies – Part 10: Carriage of timed metadata metrics of media in ISO base media file format.
 Toward a practical perceptual video quality metric, Tech blog with VMAF’s open sourcing on Github, Jun. 6, 2016.
 S.L. Regunathan, H. Wang, Y. Zhang, Y. R. Liu, D. Wolstencroft, S. Reddy, C. Stejerean, S. Gandhi, M. Chen, P. Sethi, A, Puntambekar, M. Coward, I. Katsavounidis, “Efficient measurement of quality at scale in Facebook video ecosystem”, in Applications of Digital Image Processing XLIII, vol. 11510, p. 115100J, Aug. 2020.
 R. Puri, “On a QoE metric for live media streaming applications”, Presentation in VQEG Meeting, Dec. 2020.
 R. Puri and S. Satti, “On a QoE metric for live media streaming applications”, working document, Jan. 2021.
 A. Raake, S. Borer, S. Satti, J. Gustafsson, R.R.R. Rao, S. Medagli, P. List, S. Göring, D. Lindero, W. Robitza, G. Heikkilä, S. Broom, C. Schmidmer, B. Feiten, U. Wüstenhagen, T. Wittmann, M. Obermann, R. Bitto, “Multi-model standard for bitstream-, pixel-based and hybrid video quality assessment of UHD/4K: ITU-T P.1204” , IEEE Access, vol. 8, Oct. 2020.
 G. Bjøntegaard, “Calculation of Average PSNR Differences Between RD-Curves”, Document VCEG-M33, ITU-T SG 16/Q6, 13th VCEG Meet- ing, Austin, TX, USA, Apr. 2001.
 Z. Wang, A. C. Bovik, H. R. Sheikh and E. P. Simoncelli, “Image quality assessment: from error visibility to structural similarity,” in IEEE Transactions on Image Processing, vol. 13, no. 4, pp. 600-612, April 2004.
 Z. Wang, E. P. Simoncelli and A. C. Bovik, “Multiscale structural similarity for image quality assessment,” The Thrity-Seventh Asilomar Conference on Signals, Systems & Computers, 2003, Pacific Grove, CA, USA, 2003.
 I. Katsavounidis, “VQEG’s Implementer’s Guide to Video Quality Metrics (IGVQM) project ”, working document, 2021.