Video expert assessment of high quality video for Video Assistant Referee (VAR): A comparative study

The International Football Association Board decided to introduce Video Assistant Ref-eree (VAR) in 2018. This led to the need to develop methods for quality control of the VAR-systems. This article focuses on the important aspect to evaluate the video quality. Video Quality assessment has matured in the sense that there are standardized, commercial products and established open-source solutions to measure it with objective methods. Previous research has primarily focused on the end-user quality assessment. How to assess the video in the contribution phase of the chain is less studied. The novelties of this study are two-fold: 1) The user study is specifically targeting video experts i.e., to assess the perceived quality of video professionals working with video production. 2) Six video quality models have been independently benchmarked against the user data and evaluated to show which of the models could provide the best predictions of perceived quality. The independent evaluation is important to get unbiased results as shown by the Video Quality Experts Group. An experiment was performed involving 25 video experts in which they rated the perceived quality. The video formats tested were High-Definition TV both progressive and interlaced as well as a quarters size format that was scaled down half the size in both width and height. The videos were encoded with both H.264 and Motion JPEG for the full size but only H.264 for the quarter size. Bitrates ranged from 80 Mbit/s down to 10 Mbit/s. We could see that for H.264 that the quality was overall very good but dropped somewhat for 10 Mbit/s. For Motion JPEG the quality dropped over the whole range. For the interlaced format the degradation that was based on a simple deinterlacing method did receive overall low ratings. For the quarter size three different scaling algorithms were evaluated. Lanczos performed the best and Bilinear the worst. The performance of six different video quality models were evaluated for 1080p and 1080i. The Video Quality Metric for Variable Frame Delay had the best performance for both formats, followed by Video Multimethod Assessment Fusion method and the Video Quality Metric General model.


Introduction
With the approval of the International Football Association Board to allow trials for the use Video Assistant Refereeing (VAR) technology in the game of football the Fédération Internationale de Football Association (FIFA) expressed the need to develop technical guidelines (minimum requirements) for VAR systems to approve them for the use in the game.The key questions raised by FIFA in this relation were focused on the video quality i.e., the processing of images and challenges linked to coding and decoding, the synchronisation, and the re-formatting of broadcast feeds.RISE Research Institutes of Sweden in collaboration with FIFA, therefore, developed objective test methods that can be used to measure the latency, synchronicity, and video quality of VAR-systems at three different measurement points [1], see Fig. 1.The latency and synchronisation are not covered by this article, for further information see Brunnström et al. [1].
For measurement of video quality, the following Measurement Points (MP) were identified as being important: • MP1 is where the camera signals enter the Video Operation Room and exit the VARsystem.MP2 is where the video is sent back to the Outside Broadcast van or to the broadcast provider.• Encoding, conversion of different video formats and possible integration of different image sources into single video feeds i.e., the video quality of the resulting output of the VAR system at MP1. • Measuring the output video quality from a VAR system back to the broadcaster for transmission on-air at MP2.
Video Quality assessment has matured in the sense that there are standardized, commercial products and established open-source solutions to measure video quality in an objective way [2][3][4][5].Furthermore, the methods to experimentally test and evaluate the Quality of Experience (QoE) [6,7] of a video are also widely accepted in the research community and in the broadcasting industry.These are in most cases based on standardized procedures [8][9][10][11][12][13][14][15][16][17].
Previous research and development have primarily focused on the end-user quality assessment.It is even specified in a Recommendation from the International Telecommunication Union (ITU) that a test persons should be naïve see e.g.[9].This is not so surprising Fig. 1 Schematic overview of the VAR setup.Three Measurement Points (MP) are indicated MP0, MP1, and MP2, for the evaluation of latency, synchronization, and video quality since it is where the video content is usually consumed and where most of the qualityrelated compromises must be done.However, the assessment and measurement of the video quality in the contribution phase is less studied.The European Broadcasting Union defines 'High Quality' as "a quality that is transparent to the intention of the programme maker and without visible artefacts, insofar as it is reasonably economic and technically practical to achieve."[18].
The research questions that this article is addressing are: 1) What levels of video quality are considered good or better for expert video professionals especially targeting sports content?2) Which objective quality models that have been developed by using video quality data from naïve test persons and most likely a wider and lower quality range, can well predict the quality in a high-quality range as defined by expert professionals from the broadcast industry?
The novelty and the aim of this research and paper is twofold: • Firstly, a user study was conducted specifically targeting video experts.That is using professionals whose main occupation is video production.The study measured how they perceived the quality of the shown videos.This gives valuable insight into the video quality needed in the contribution phase.• Secondly, the objective quality methods performance in predicting video experts' opinions have not been performed previously.In this article, we are not presenting our own method in comparison to others.This is to be able to be unbiased and present an impartial evaluation.There is a tendency to publish the creators' solution as the best performing method, with usually inflated prediction performances and it is unclear whether all methods that have been compared to had received the same chance.In independent evaluations, the performances usually drop drastically [19], compared to the performances in the articles where the methods were originally presented.This has also been shown repeatedly by the work of the Video Quality Experts Groups (VQEG) [20][21][22][23][24][25][26][27].
Video quality models are objective, mathematical models that approximate results from user quality assessments, in which human observers are asked to rate the quality of a video.In the literature various models and approaches are suggested [28,29].They are very often subdivided into groups depending on the amount of reference information that the models are using.
• Full Reference (FR) -access to high quality or non-distorted versions of the video to compare with the distorted video.• Reduced Reference (RR) -key quality parameters are computed from the reference and the distorted video, which are compared [30].• No Reference (NR) -no access to the reference [30].
Many models have been proposed in the literature, e.g., [31][32][33][34].The Video Quality Experts Group (VQEG) has been involved in several efforts to evaluate the performance of models proposed [20,21,[23][24][25]29], which later has been standardized by the ITU [29,31,33].In this research, we primarily focused on the FR models as they are more accurate than the NR models.Additionally, FR models are more likely to cover the high-quality range that is required for the use in VAR systems [35,36].A complication is that it requires a reference signal i.e., a video sequence taken at MP 0, that can be compared to the same video sequence at MP 1 and MP 2. For a working solution see [1].In this article, we have chosen to present the performance of open source and published models.The performance of some commercial products were also investigated.As they did not perform better than the best-performing models in our study, we have made the decision not to disclose their names and performances.
The evaluation of the objective models in this work is based on data from a user study performed using standardized procedure as recommended by the ITU [8][9][10].The performance comparison between the objective quality models was also following recommendations by the ITU [37].
The rest of the article is organized starting with on overview of related work in Section 2. The method used is then described in Section 3, followed by the Results in Section 4. The article is concluded with a Discussions and Conclusions in Sections 5 and 6 respectively.The abbreviations used are summarised in Table 1 and the symbols in Table 2.

Related work
Fully independent evaluations of objective video quality methods are rare.Many of the independent evaluations that have been conducted were done by VQEG and recently also by the ITU-T Study Group 12, either independent or in collaboration with VQEG.For the work of VQEG we will cover only those evaluations that most resembles the current work.Please, see www.vqeg.org for a full list of evaluations performed.
In VQEG phase I [25] objective video quality methods were evaluated for standard definition TV.It was a large test where about 26000 user opinions were collected, that was divided into four different parts 60 Hz high and low-quality and 50 Hz high and low quality.The high-quality range was with bitrates between 3 Mb/s and 50 Mb/s, whereas the low-quality range had bitrates between 768 kb/s and 4.5 Mb/s.There were nine methods evaluated.The effort was at the time seen as a failure as no method proved to be statistically better than Peak Signal to Noise Ratio (PSNR) [38].The high-quality range of this work was similar to the current work, although with another video format Standard Definition TV (SDTV) as compared to High Definition TV (HDTV) in this study.The experimental method was also different as in VQEG phase I the Double Stimulus Continuous Quality Scale (DSCQS) [8] was used as compared to Absolute Category Rating (ACR) with Hidden Reference (ACR-HR) [9] removal that was used in this work.Furthermore, naïve observers were used in VQEG phase I and not video experts as in this article.The work was still very important as it showed very clearly for the first time the drop in performance of the video quality methods when they were subjected to a fully independent evaluation.VQEG phase II [24] was not as large as phase I but was designed to be more challenging for the metrics.It did not contain the high-quality range of the videos.The experimental method was DSCQS and the observers were naïve.
VQEG HDTV [20] is in a sense the experiments most similar to current investigation.The video formats were 1080p at 25 and 29.97 frames-per-second and 1080i at 50 and 59.94 fields-per second.The encoding considered were MPEG 2 and H.264 with both encoding distortions and transmission errors.The bitrates range were between 1 -30 Mbit/s.The experimental method for getting the user rating was the same is in the current study i.e., ACR-HR.The observers were naïve as a contrast to the current study.
Wulf and Zölzer [39] evaluated three different objective quality metrics on three open video quality databases but did not perform their own user study.The majority of The ratio between the maximum RMSE squared and the minimum RMSE squared the data comes from VQEG HDTV and the rest from Ecole Polytechnique Fédérale de Lausanne and Politecnico di Milano [40,41] as well as Laboratory for Image and Video Engineering [42].Sedano et al. [43] used the same databases as Wulf and Zölzer and evaluated six different full-reference methods for the purpose of developing a light-weight parametric noreference model.The paper presented a method of using full-reference methods that was tested for a specific scope for the development of no-reference modules for that scope.
Raake et al. [29] did a comprehensive evaluation of Bitstream-, Pixel-Based and Hybrid Video Quality methods for the development of three standards for video quality assessment for video formats up to 4 K resolution involving video codecs such as Advanced Video Coding (AVC) (H.264) [44], H.265/HEVC [45] and VP9.The user studies were using ACR as in the current study, but the ratings were based on naïve observers.
Ling et al. [46] performed a study for high quality video involving both HD and UHD videos.The user studies used the Degradation Category Rating (DCR) method instead of ACR.Four different objective quality methods were independently benchmarked in the study.The viewers were non-experts.
Pinson [19] did a comprehensive independent evaluation of no-reference models that had been published in the literature.The general conclusion was that the performance that had been published by the authors of the models did not come near the performance when they were independently tested.The main differences here are the focus on no-reference models and video quality databases based on naïve observers.
In summary (see Table 3) previous studies have been based on using non-expert or naïve observers in the investigations.This is not surprising since this is a common practice and recommended by the ITU.In our case, the videos used were close in quality to what could be expected in the contribution phase in broadcasting, the appropriate observer should be an expert in judging video quality.

Procedure
For high video quality where the differences are small, there are arguments for a comparative methods like the Double Stimulus Impairment Scale (DSIS) or Double-stimulus continuous quality-scale (DSCQS) [8].However, these will reduce the number of stimuli or videos that can be scored during an experiment, since each trial takes either almost double the time for DSIS or up to four times as long when using DSCQS [8], as compared to a single stimulus method.It is also important to get a wide variety of video rated to have a rich data to evaluate objective methods against.We, therefore, decided to use a single stimulus method: ACR-HR method [8][9][10].It was shown by Lee et al. that the results correlate well with DSCQS [11].This approach has been successfully employed previously e.g., VQEG HDTV [20].
This method uses single stimulus procedure.One video is presented at a time to the user and they are asked to provide their rating on a five graded category scale for each video after the video stops, see Fig. 2 The ratings were provided in this study via a voting interface on the screen, see Fig. 3, asking the user to "judge the video quality of the video?"(In  Swedish: "Bedöm videokvaliteten hos videon?").The rating scale used was the five graded ACR quality scale [8,9] (Swedish language translation in parentheses): 5 Excellent (Utmärkt) 4 Good (Bra) 3 Fair (Varken Bra eller Dålig) 2 Poor (Dålig) 1 Bad (Usel) High-quality or pristine reference videos were mixed in among all the other videos and rated as the others.Viewers saw each presentation of a video once and did not have the option of re-playing a presentation.However, they could change their voting as many times as they wanted before submitting their final answer, as described in ITU-T P.913 [10].
To evaluate objective video quality models for different formats that are used in the TV production of football games the users were asked to provide their rating on three different video formats: 1) Full size 1920 × 1080 video based on progressive source (1080p).
The order in which the different video formats were played to the users, as well as the order of the single video sequences within each video format were randomized for each participant.For the video playback and randomisation, the VQEGPlayer [32] was used.The time required by the users to watch the videos and provide the rating for all three sessions was in total approximately 45 min, with short breaks between each session.The total time required for each user, including instructions, visual testing, training, pre-, and post-questionnaire, was about 1.5 h.
There were 60 Processed Video Sequences (PVS) to be evaluated per session.These consisted of 6 different SouRCe sequences (SRC), i.e., different content that each of them was processed with 10 different error conditions, so called Hypothetical Reference Circuits (HRC), see also below in Sessions 3.1.4and 3.1.50.Each video was 10 s and with an average estimated voting time of 5 s, a trial was about 15 s.This gave a total time for a session of approximately 15 min.
Instructions were written out for the participants to read, to ensure that the instructions given were as similar as possible.Some explanations and backgrounds were given verbally, especially in response to any questions and uncertainty of the task to perform.Before the actual test a training session was carried out to familiarize the test persons with the procedure and to show the range of qualities involved.A very good, an intermediate and a bad quality version of a video that was not used in the actual test were shown to the test persons.Test persons were allowed to ask questions after the training session if something was still unclear.

Set-up of test room
To create a controlled and uniform environment for the participants the test room was setup to comply with the requirements of the ITU-R Rec.BT.500-15 [8].A high-end consumer-grade 65″ 4 K TV (Ultra HD, LG OLED65E7V) was used for the experiments, having a resolution of 3840 × 2160 pixels.As the videos used in the experiment had a lower resolution (1920 × 1080 and 960 × 540) than the screen, the video was displayed pixel matched in the centre of the screen with a grey surround, see Fig. 4. The interlaced 1080i video was deinterlaced in software (see Section 3.1.5)and the deinterlacing of the TV was not used.The TV was characterized e.g., luminance, contrast, colours, and colour temp, and verified to conform to ITU-R BT.709 [47].
Different video formats and display resolution have been optimized to be viewed from a specific viewing distance, roughly corresponding to an angular pixel pitch of 1 min of arc.For HD (1920 × 1080) this is 3H, where H = Picture Height (height of the video window, not the physical display), corresponding to 120 cm in our case.The users were requested to keep this viewing distance while evaluating the videos.This distance was marked on the floor and the chair, which was an office chair on wheels that had been equipped with a wheel locking mechanism.The position of each participant was also adjusted, so that their eyes were vertically and horizontally centred in relation to the TV screen centre when looking straight ahead.

Test persons
In the experiment, 25 Swedish-speaking video experts participated, see further Section 4.1.1.
The term 'expert' has been used in the sense that the viewers' work involves video picture quality, picture production in broadcasting, creation of film or video, post-production, etc.They were recruited by an external company Feebackfrog AB (www.feedb ackfr og.com), which has been used previously by RISE for recruitment of participants in Usability testing.Feedbackfrog took all the contacts with the participants, scheduled the time Fig. 4 The video with resolution 1920 × 1080 pixels was presented in the centre of the screen with a grey surround to experiment, and, finally, also expedited the compensation after the completion of the experiment.The anonymity of the data collected was, therefore, ensured as no contact information of the participants were digitally stored at RISE.The participants were compensated with gift cards to a value of 1500 SEK (about 150 EUR) each.
All viewers were tested prior for the following: • Visual acuity with or without corrective glasses (Snellen test).

Source video sequences
To rate the video quality a set of six different source video sequences was shown to the expert panel.The video formats selected for the SRC were: • 1920 × 1080 progressive 50 frames-per-seconds (1080p) • 1920 × 1080 interlaced 50 fields-per-seconds (1080i) All SRC (Tables 4 and 5) were obtained as uncompressed videos during live football broadcast productions, as well as from Swedish Television's, "Sveriges Television (SVT)", production Fairytale that was produced for research and standardization purposes [48].From all collected videos, video clips of the length of 14 s were extracted.
The collection of the football related videos took place in Madrid, at the Estádio Municipal de Butarque on 30 April 2019, during a FIFA test event for Virtual Offside Line Technology where both video formats, 1080p and 1080i, were recorded.The second football video was collected during a match of the Swedish national football League at the Tele2 Arena in Stockholm on 14 th of May 2019.Due to the broadcasting of the football match that was conducted in 1080i format only this video format was collected.In both cases the equipment to collect the videos was a Blackmagic Design UltraStudio HD Mini grabbing box connected to a Lenovo Carbon X1 6 th gen computer (CPU: Intel Core i7-8550U, RAM: 16 GB) and an external 500 GB SSD hard drive (Samsung X5) with Thunderbolt connection.The software used was Blackmagic Design Media Express Version 3.5.7 and the format was uncompressed AVI 16 bit 4.2.2 uyvy.
The SRC produced from the SVT production are clips from a 6.5-min-long video Fairytale [48], that was professionally filmed and produced on 65 mm analogue film in 50 fps (slow motion up to 100 fps) and scanned frame by frame while colour correcting and applying film grain noise reduction, to produce the 4 K (3840 × 2160 progressive, 16 bit per colour) Master.The 1080p version was produced by downsampling the Master using a sinc filter.
The interlacing was also carefully done, by starting with a 2164-line (3840 × 2164p/50) raster.Every second frame was shifted two lines downwards and then they were cropped to 2160 lines.The shifted half frames were filtered down to 540 each, by Shake's box filter vertically and Shake's sinc filter horizontally to reduce the resolution to 1920 [49,50].For more details on the production see [48].
When selecting the SRC the focus was primarily on football content.To ensure diversity in the video sequences the content was mixed with the Fairytale content.To be representative for the future use of the test method, in the analysis of the video quality of VAR Technology, a special requirement for the videos were that they had to contain overall a lot of motion to be representative for football.
Final SRC were cut to 10 s, but the source scenes used for PVS creation were 14 s to accommodate for any transient behaviour of the encoder.

SVT_ParkJoy (SRC6)
A few people running at a distance, the camera is moving with them, and trees are passing in the foreground.High contrasts.This sequence is defined in [48] and classified as "difficult" on the coding difficulty.Grabbed at Tele2 Arena.

Processed video sequence generation
A Hypothetical Reference Circuit (HRC) is a processing performed by a system or software that can introduce degradations to the video e.g., video encoding, scaling, transmission, etc. [10].The processing performed in this experiment has been similar for the different formats that were used in the experiment, but with differences adjusted to the format as well as targeting specific aspects of the format.There were 10 different HRCs per video format (including the reference) and each HRC was applied to each SRC for each of the formats, making 60 PVSs per format [10].All PVSs was 10 s long.
A summary of the HRCs is the following:

Table 8
The HRCs or processing applied to the 540i format The experimental program VQEGPlayer [51], takes the raw format Blue, Green and Red (BGR) and not AVI as a file format, and a colour space conversion was, therefore, applied to PVS:s as a the last step.In addition, the interlaced video clips (1080i and 540i) were deinterlaced in software.
• FFMPEG -vcodecrawvideo -pix_fmt uyvy422 -vfyadif = 1:0:0,mcdeint = 3:0:1 The last step in the preparation of PVS was performed by the program iconvert, written by Prof. Marcus Barkowsky, Deggendorf Institute of Technology (DIT), Germany.The program performed format and colour space conversion, as well as cutting off the 2 first and 2 last seconds of the PVS down to 10 s.

Objective video quality assessment methods
In the evaluation we were interested to study how well the models could predict the data obtained from the users.We have followed standardised procedures from the ITU in this regard [37].We studied the overall performance by calculating the Pearson Correlation Coefficient (PCC) [37] and the Root Mean Square Error (RMSE) [37], between the user data Difference Mean Opinion Scores (DMOS) and the scores of the objective models i.e. the Predicted DMOS (PDMOS).The PCC measures the linear relationships between the DMOS and the PDMOS.The RMSE gives the accuracy.
The DMOS was calculated by subtracting for each user its rating of the reference from the rating of the distorted video.To get the values on the same scale as the Mean Opinion Scores (MOS) i.e., 1-5, the following formula was used: Here the index k was running over all SRCs (P = 6) and i over all HRCs (Q = 10) and j over all participants (N = 25).The reference for source k was denoted SRC k = PVS k1j, which was then the PVS for the HRC with index 1.
(1) The PCC measures the linear relationship between the model scores and the DMOS.For simplicity we drop the index to keep track of the different sources and let the index i be taken over all PVSs i.e., i = 1 … M .where M = 6 ⋅ 9 = 54.
where DMOS and PDMOS were the mean values of respective entity.The calculation was done using the Matlab function corr [53].The RMSE was calculated using: where M is the number of points that are compared and minus 1 is to get an unbiased estimate of the RMSE according to ITU-T Rec.P. 1401 [37].
As the relationships very often are not linear it is recommended to linearize the dependency by fitting a 3 rd order monotonic polynomial to the data [37].We also use this to map the value range of the PDMOS into the range of the DMOS i.e., 1 to 5 so that RMSE can be calculated.According to [37] we have used to following mapping: where RMSE(DMOS, PDMOS) → min and f PDMOS raw = monotonous between PDMOS min and PDMOS max .
Finally, a statistical hypothesis test was applied to the RMSE values.The null hypothesis, H 0 , was that there was no statistical difference between two RMSE values, and the alternative hypothesis, H 1 , was that there was a statistical difference.The test was based on forming an F ratio between the larger RMSE value squared divided with the smaller RMSE value squared.The degrees of freedom was the number of points in the RMSE calculation, minus 4 due to the 3 rd order monotonic polynomial fit i.e. 54 -4 = 50 [37].
The q-value was then compared to F(0.05, 50, 50) for 95% confidence and if q > F the difference was considered significant.We used Holm [54] to handle multiple comparison.The Spearman Correlation Coefficient (SCC) was also calculated [55], using Matlab function corr [53].
Another criterion for a useful method is whether it can handle registering the distorted video with the reference.That is if some offset has occurred either spatially or temporally by the distortion introduced.This can have a severe impact on some computations.For instance, the Peak Signal to Noise Ratio (PSNR) computes the difference pixel by pixel and a small shift can, therefore, have a big impact on the score, which is not reflected by a quality difference.The Video Quality Metric (VQM) [31] software also contains a rather sophisticated registration algorithm that could be run separately from the core algorithm and could, therefore, be used for pre-processing of videos to evaluate the performance of other algorithms.This registration method has been standardized by the ITU (ITU-T Rec J.144) [5].

Video Multimethod Assessment Fusion
Video Multimethod Assessment Fusion (VMAF) [3] was developed by Netflix and the University of Southern California.It is not standardised but has shown good performance in recent evaluations.The method is available as open-source software (https:// github.com/ Netfl ix/ vmaf/ relea ses).VMAF fuses the results from more basic image qualitybased metrics, but also from motion estimation calculations.The fusing was done using a Support Vector Machine (SVM) regressor [56].In [3] presents three metric used: • Visual Information Fidelity (VIF) -see below.In VMAF the metrics output from different scales are kept separate and not combined over four scales as in the original formulation.• Detail Loss Metric (DLM) [57] -this is an algorithm that measures loss of useful information and added impairments that will distract the user.In the original algorithms a combined score is produced, but in VMAF the elementary metrics are used.• Motion is temporal estimation based on the average absolute pixel difference of the luminance component between consecutive frames.

Video Quality Metric
The Video Quality Metric (VQM) was developed by the Institute for Telecommunication Sciences (ITS), the research laboratory of the National Telecommunications and Information Administration (NTIA) (USA).It was standardized by the ITU for SDTV (ITU-T Rec.J.144) [5], and has shown very good performance, especially the updated version VQM for Variable Frame Delay (VQM_VFD) [2].It is available as open-source software.
The VQM General model was designed to be a reduced reference model but has in most cases been compared to FR-models and the performance has been as good or even better.It was designed for a wide variety of video systems and bitrates.VQM was also tested and evaluated extensively during its development phase [31].The model was developed and optimized for the codecs and video system of its time i.e., before 2004.After performing a rigorous calibration that will contain spatial and temporal alignments, calculations of valid region as well as gain and level offset estimation, the video quality is estimated as a linear function of some quality features.where si_loss is the loss of spatial information such as blur; hv_loss is the degree of that the horizontal and vertical edges have shifted towards a more diagonal orientation; hv_ gain detects the opposite shift of edges compared to hv_loss; chroma_spread computes the changes in the spread of 2D colour samples; si_gain measures quality improvements based on e.g.edge sharpening and chroma-extreme finds colour impairments due to transmission errors.The command used for VQM General (version 30) was: matlab.exe-r cvqm ('reference.avi','distorted.avi','progressive', 'frcal', 'general').
VQM_VFD was designed to cope with videos that can have dropped frames or in other ways have variations in their frame rate.The key difference with this updated model is the neural networks module weighting together the extracted perceptual features of the model [2], as shown in Equation above.

Peak Signal to Noise Ratio
Peak Signal to Noise Ratio (PSNR) is the most used video quality assessment method.It is standardized by ITU in ITU-T Rec J.340 [38].It has been disproven many times for not providing a good estimate of perceived visual quality.It has merits in the high-quality range, but its output values cannot be compared between different source content, which makes it hard to attach requirement levels to that the system should fulfil.
Open source implementations are provided by Ecole Polytechnique Fédérale de Lausanne with VQMT: Video Quality Measurement Tool (VQMT) [58] and within the VQM software [59].
This implementation calculates PSNR frame by frame by first forming the Mean Square Error of the difference between the reference and the distorted frame The calculations were performed in the colour space YUV 4:2:2.The different channels carried equal weights in the calculations.The command used for calculating PSNR was: vqmt.exereference_video.yuvdistorted_video.yuv1080 1920 500 yuv422 results.txtPSNR.
The overall result was obtained by taking the average over all frames.

Structural Similarity Index
The Structural Similarity Index (SSIM) is an image quality method that has been popular since it was first published [32].It was developed to improve compared to PSNR but is still very simple.The idea is to compare the structure between two images.It is defined for images and does not specify how it should be used for video.The paper [32] is one of the most cited papers in the image processing community.Here we have used the implementation in VQMT [58].A usual form to calculate the SSIM is where x and y represent the different images that are compared, µ is the mean and σ standard deviation.C 1 and C 2 are constants to avoid instability in the calculations.The command used was the same as above for PSNR and calculated at the same time and the colour space was treated the same.The overall result was obtained in the same way by taking the average over all frames.

Visual Information Fidelity
Visual Information Fidelity model has an information-theoretic approach to the image and video quality assessment problem.The model combines an image fidelity criterion of shared image information between the reference and the degraded image, with a measure of the image information that is present in the reference and calculates how much of it that can be extracted from the degraded image [34].Here we have used the implementation in VQMT [58], which implements the pixel domain version.
In the VIF method the image is divided into subbands for instance with the discrete wavelet transform [60,61] or using the method of steerable pyramid [62].A VIF metric is calculated for each subband using where ref for the subband the average values of the luminance of the reference and dist the same for the distorted image.ref ,dist is the covariance between the reference and the distorted image.ref and dist are the standard deviations for the respective images for the subbands.The total VIF is obtained by multiplying all the subband VIFs together.For N subbands The command used was the same as above for PSNR and calculated at the same time and the colour space was treated the same.The overall result was obtained in the same way by taking the average over all frames.

Characterisation of expert panel
In total 25 video experts participated: 23 males and 2 females.The number of females was low, but this was expected considering the area of investigation.The average age of the test users was 37.8 years, with a standard deviation of 10 years.The oldest was 65 years of age and the youngest was 24 years.All participants were working in or studying in the TV area as producers, technicians, and photographers, or similar, thus considered as experts in evaluating video quality.This was shown in their answer to the question of their experience of video evaluation where 1 meant "no experience" and 5 meant "expert".This question had an average of 4.5 and a standard deviation of 0.58 and only two users scored below 4 one put 3.5 and the other 3. ( 11) Very few had the experience of being scientific video test viewers.The average was as low as 1.48 (standard deviation 0.75) where 1 meant no experience and 5 a lot.Only one subject put a 4 and one put a 3.About 1/3 of the subjects put 1.
All test persons had a good visual acuity as expected for such professionals, average 1.09/1.06(right/left eye), standard deviation 0.18/0.20,max 1.4, and min 0.6 on one eye.About half of them wore glasses or lenses.All had an accurate colour vision.
The test persons were asked to rate how typical they experienced the distortions in the videos from 1 to 5, where 1 was not at all and 5 was very typical.The mean was 4 with a standard deviation of 1, showing that the test persons were very familiar with the distortions.
The test persons were also asked if the range of resolutions was typical on the same scale as above and mean was 3.4 with standard deviation 1.3.This indicates that the range was not completely typical, but more typical than not.

Statistical analysis of users' rating data
The scale responses were given numerical values when analysed using the following: Bad = 1, Poor = 2, Fair = 3, Good = 4 and Excellent = 5.
Characterization of the quality of the video clips is the Mean Opinion Scores (MOS) which is the mean over the ratings given by the users where ij is the score of user i for PVS j.N is the number of users and M is the number of PVSs.
The statistical analysis that was performed was by first applying a repeated measures Analysis of Variance (ANOVA) and then performing a post-hoc analysis based on Tukey Honestly Significant Difference (HSD) [63,64].The repeated measures ANOVA will determine whether there are any significant differences between the MOS, and the post-hoc test will identify the specific pairs for which these significant differences exist.
Another characterisation of the data is the Standard deviation of Opinion Scores (SOS) as suggested by Hossfeld, et al. (2011) [65], which is now widely used.They demonstrated that there is a squared relationship between the SOS and MOS, dependent on only a single parameter a, which can be used to characterize an experimental study.For instance, they showed that typical image and video quality experiments has an a between 0.15, whereas a web surfing experiment would be about 0.25 and gaming above 0.3.The objective of this method is to determine the a value that provides the best fit to the data.

1080p
The mean quality of the source video clips (SRCs) for 1080p taken over all degradations (HRCs) and users is shown in Fig. 5 left graph (see also Table 4).The MOS values are close to 3 and slightly above indicating that they have about the same number of highquality degradations as low degradation with slightly more high quality than low quality.SRC2 (Butarque_follow_player_2) had the overall highest quality, indicating that this video was the least challenging to encode.SRC4 (SVT_CrowdRun) had the overall lowest quality, indicating that this video was the most challenging to encode, followed by SRC6 ( 13) (SVT_ParkRun).These two had also statistically significantly lower quality than SRC2, having p-values lower than 0.05 (p = 0.00025 and p = 0.038 respectively).
In Fig. 5 right graph (see also Table 6) the mean quality of the degradations (HRCs) taken over all source video clips (SRCs) and users are shown.It can be noted that there are some HRCs (4-7) that have almost the same quality (MOS) as the reference (HRC1).These are not statistically significantly different from the reference.The other HRCs were statistically significantly different from the HRC1 with p = 0.00001 < 0.05.
A breakdown of the HRCs into the different processing schemes and bitrates that were applied to the SRCs is shown in Fig. 6.The encoding performed by Motion JPEG is shown in solid black and the H.264 in dashed black curve.The MOS of the reference is marked as a red line without tying it to the bitrate to not make the x-axis too long.The quality drops fast with lower bitrates for MJPEG, whereas the quality for H.264 is indistinguishable from the reference down to about 20 Mbit/s.
In Fig. 7 the relationship between SOS 2 and MOS along with the fitted curve that estimates the parameter a for 1080p is shown.The parameter was estimated to be 0.19.Here Fig. 5 (left) Mean quality of the source video for 1080p (y-axis) clips (SRCs, x-axis) for 1080p taken over all degradations (HRCs) and users, see also Table 4. (right) Mean quality (y-axis) of the degradations (HRCs, x-axis) taken over all source video clips (SRCs) and subjects, see also Table 6.The error bars represent 95% confidence intervals Fig. 6 The mean quality for 1080p (y-axis) of the degradations (HRCs) taken over all source video clips (SRCs) and users, divided into the different codecs used (MJPEG in solid black curve and H.264 dashed black curve) and plotted against the bitrate (x-axis).The MOS of the reference is marked as a red line without tying it to the bitrate to not make the x-axis too long.The error bars represent 95% confidence intervals we can also see that there more videos rated with higher quality than with lower quality.The variance is also highest about 3 to 3.5 showing that in this range the test persons were most uncertain about the quality.

1080i
In the same way as for 1080p above, the mean quality of the source video clips (SRCs) 1080i taken over all degradations (HRCs) and users is shown in Fig. 8 left graph (see also Table 5).Fig. 7 The relationship between SOS 2 and MOS along with the fitted curve that estimates the parameter a for 1080p Fig. 8 (left) Mean quality for 1080i (y-axis) of the source video clips (SRCs, x-axis) for 1080i taken over all degradations (HRCs) and users, see also Table 5. (right) Mean quality (y-axis) of the degradations (HRCs, x-axis) taken over all source video clips (SRCs) and users, see also Table 7.The error bars represent 95% confidence intervals The MOS values are also in this case slightly above 3. SRC5 (SVT_CrowdRun) and 6 (SVT_ParkJoy) have statistically significantly lower (p < 0.05) MOS than SRC1 (Football_celebrating_1, p = 0.0001).SRC3 (Football_greetings_1, p = 0.0001) and SRC4 (Football_greetings_1, p = 0.0007) and SRC2 (Football_goal_1) have lower MOS than SRC1 (p = 0.005) and SRC3 (p = 0.02).
In Fig. 8 right graph (see also Table 7) the mean quality of the degradations (HRCs) taken over all source video clips (SRCs) and users are shown.Also, for 1080i there are some HRCs (3)(4)(5) that have almost the same quality (MOS) as the reference (HRC1).These are not statistically significantly different from the reference.However, HRC6 is statistically significantly lower, even if this can be hard to see from the graph (p = 0.03).The other HRCs, were also statistically significantly different from the HRC1 with p = 0.00001 < 0.05.
The MOS plotted against the bitrates for the different HRCs are shown in Fig. 9.The encoding performed by MJPEG is shown in solid black and the H.264 in dashed black curve.The MOS of the reference is marked as a red line without tying it to the bitrate to not make the x-axis too long.The HRC10 was not an encoding error condition and was just a very simple deinterlacing applied directly to the uncompressed video and its MOS has been drawn in a similar way as the reference HRC1, as a yellow line across the graph.This error condition was not liked very much by the users and thus received very low ratings.The quality drops fast with lower bitrates for MJPEG, whereas the quality for H.264 is indistinguishable from the reference down to about 30 Mbit/s, but in contrast to 1080p 20 Mbit/s is statistically significantly lower for 1080i (p = 0.03 < 0.05).
In Fig. 10 the relationship between SOS 2 and MOS along with the fitted curve that estimates the parameter a for 1080i is shown.The parameter was estimated to 0.18, almost the same as for 1080p.Here we can also see that there are more videos rated with higher quality than with lower quality and even more compared to 1080p.The variance was in this case highest in the range 3.5 to 4 further indicating a more skewed distribution than 1080p.
Fig. 9 The mean quality for 1080i (y-axis) of the degradations (HRCs) taken over all source video clips (SRCs) and users, divided into the different codecs used (MJPEG in solid black curve and H.264 dashed black curve) and plotted against the bitrate (x-axis).The MOS of the reference is marked as a red line without tying it to the bitrate to not make the x-axis too long.Similarly, the error conditions based on simple deinterlacing on an otherwise uncompressed video are shown as a yellow line.The error bars represent 95% confidence intervals

540i
The mean quality of the source video clips (SRCs) for 540i taken over all degradations (HRCs) and users is shown in Fig. 11 left graph (see also Table 5).The average quality i.e., the MOS values are here more above 3 than what could be observed for 1080p (Fig. 5) and 1080i (Fig. 8).This is most likely due to the smaller format, which makes it harder to see the degradations.Overall highest quality had SRC3 (Football_greetings_1), which had a statistically significantly higher MOS than all the other SRCs (p < 0.05).The MOS of SRC6 (SVT_ParkJoy) was statistically significantly lower (p < 0.05) than SRC5 (SVT_ CrowdRun, p = 0.0003) and SRC4 (Football_play_1, p = 0.0002) as well.
Fig. 10 The relationship between SOS 2 and MOS along with the fitted curve that estimates the parameter a for 1080i Fig. 11 (left) Mean quality (y-axis) of the source video clips (SRCs, x-axis) for 540i taken over all degradations (HRCs) and users, see also Table 5. (right) Mean quality (y-axis) of the degradations (HRCs, x-axis) taken over all source video clips (SRCs) and users, see also Table 8.The error bars represent 95% confidence intervals In Fig. 11 right graph (see also Table 8) the mean quality of the degradations (HRCs) taken over all source video clips (SRCs) and users are shown.The different bitrates applied with H.264 come in groups of three, so HRC2-4 is 50 Mbit/s, HRC5-7 is 20 Mbit/s and HRC8-10 is 10 Mbit/s.The latter group was statistically significantly lower quality than the other, including the reference HRC1 (p < 0.05).For the other groups, it depends on which scaling method was used whether there was a difference or not.HRC6 bilinear 20 Mbit/s was statistically significantly worse than 50 Mbit/s (p < 0.05), but not Lanczos HRC5 and nearest neighbur HRC7.HRC3 bilinear 50 Mbit/s was statistically significantly worse than the reference HRC1 and Lanczos 50 Mbit/s HRC2.Lanczos were statistically significantly better than bilinear for all the bitrates.
A breakdown of the HRCs into the different processing schemes and bitrates applied to the SRCs is shown in Fig. 12.The different scaling methods are drawn as a separate curve, where Lanczos is drawn in solid black, bilinear in dashed black, and nearest neighbour in yellow.The MOS of the reference is marked as a red line without tying it to the bitrate to not make the x-axis too long.The graph shows a clear drop in quality for 10 Mbit/s, but not so severe for 20 Mbit/s (both levels were significant with p < 0.001), and hardly any quality decrease for 50 Mbit/s.The difference in quality between Lanczos and bilinear were statistically significant for 50 Mbit/s (p = 0.003), 20 Mbit/s (p = 0.0002) and 10 Mbit/s (p = 0.0005).
In Fig. 13 the relationship between SOS 2 and MOS along with the fitted curve that estimates the parameter a for 540i is shown.The parameter was estimated to 0.22, which is higher than both 1080p and 1080i.The distribution of qualities was more evenly spread than for the other formats.The variance is also highest approximately 3 showing that in this range the test persons were mostly uncertain about the quality.
Fig. 12 The mean quality (y-axis) of the degradations (HRCs) taken over all source video clips (SRCs) and users, divided into the different scaling methods (Lanczos in a solid black curve, bilinear in dashed black curve, and nearest neighbour in yellow curve).The scaling was combined with H.264 encoding.The MOS of the reference is marked as a red line without tying it to the bitrate to not make the x-axis too long.The error bars represent 95% confidence intervals

Objective video quality models evaluation
The objective models presented in Section 3.2, were evaluated for their performance on the video format 1080p and 1080i, using the method described there.

Video Multimethod Assessment Fusion
The performance of VMAF was very good using the default training.The results was a PCC of 0.89 and RMSE of 0.44 for 1080p (see Fig. 14 and Table 10) and a PCC of 0.89 Fig. 13 The relationship between SOS 2 and MOS along with the fitted curve that estimates the parameter a for 540i Fig. 14 Scatterplots of VMAF performance for 1080p.The left graph shows the results after the 3 rd order monotonic polynomial fit with a linear correlation line overlaid.The right graph shows the results before the 3 rd order monotonic polynomial fit with the fitted function overlaid and RMSE of 0.47 for 1080i (see Fig. 15 and Table 11), based on the 3 rd order monotonic polynomial fit which (see Tables 14 and 15).Scatterplots are shown in Fig. 15.This implementation of VMAF needs external registration software, i.e., the encoding may shift the reference video and the degraded video a few frames which needs to be adjusted before running the model otherwise very low scores are obtained.

Video Quality Metric
VQM General model (ITU-T Rec.J.144) was evaluated at RISE and the performance was good, with a PCC of -0.87 and RMSE of 0.49 for 1080p (see Fig. 16 and Table 10) and a PCC of -0.86 for 1080i and RMSE of 0.51 (see Fig. 17 and Table 11), based on the 3 rd order monotonic polynomial fit (see Tables 14 and 15).Negative values indicate that the slope of the line is negative, meaning in this case if the quality is low the model gives higher values than if the quality is high.VQM has also an improved model, called VQM_VFD, where VFD stands for Variable Frame Delay.RISE evaluated this version too, which had a really good performance and was the top-performing model in this evaluation, with a PCC of -0.96 and RMSE of 0.26 for 1080p (see Fig. 18 and Table 10) and a PCC of -0.91and RMSE of 0.41for 1080i, (see Fig. 19 and Table 11), based on the 3 rd order monotonic polynomial fit (see Tables 14 and 15).

Peak Signal to Noise Ratio
PSNR was evaluated with a PCC of -0.75 and RMSE of 0.65 for 1080p (see Fig. 20 and Table 10) and a PCC of -0.81 and RMSE of 0.60 for 1080i, (see Fig. 21 and Table 11), based on the 3 rd order monotonic polynomial fit (see Tables 14 and 15).In the raw format (Fig. 21 right graph), there is a group of outliers belonging to HRC10 (simple deinterlacing), that decrease the performance.The 3rd-order monotonic polynomial fit partly compensated for it and included them into the main group.
The implementation of PSNR needs external registration software but could also be calculated with the VQM software and then the registration could be done using that [38].

Structural Similarity Index
SSIM performance was evaluated to a PCC of -0.77 and RMSE of 0.63 for 1080p (see Fig. 22 and Table 10) and a PCC of -0.66 and RMSE of 0.76 for 1080i, (see Fig. 23 and Table 11), based on the 3 rd order monotonic polynomial fit (see Tables 14 and 15).SSIM has an especially low performance for 1080i.This is due to the HRC10 which SSIM had large problems with, and the 3 rd order monotonic polynomial fit was not able to adequately compensate for it.
This implementation of SSIM needs external registration software.

Visual Information Fidelity
VIF performance was evaluated to a PCC of -0.75 and RMSE of 0.64 for 1080p (see Fig. 24 and Table 10) and a PCC of -0.80 and RMSE of 0.61 for 1080i, (see Fig. 25 and Table 11), based on the 3 rd order monotonic polynomial fit (see Tables 14 and 15).This implementation of VIF needs external registration software.VIF has similar problems with HRC10 as PSNR and SSIM.

Statistical significance testing
The results of statistical hypothesis testing, based on a pairwise comparison of the RMSE values, see Section 3.2, for a description of the procedure that is based on ITU-T Rec.P.1401 [37].The p-values are shown in Tables 12 and 13.VQM_VFD is significantly better than all other models for 1080p and better than PSNR, SSIM, and VIF for 1080i.VMAF is significantly better than PSNR and VIF for 1080p.SSIM has a very low performance for 1080i and is significantly worse than all other models (Tables 14 and 15).

Discussions
This study was motivated by the need to find a video quality algorithm that is good in predicting quality for videos that are presented for assistant referees.Typically, the bandwidth used and thus the quality is higher than what is normally provided for end TV consumers.Therefore, video experts were invited to the experiment for collecting relevant data to differentiate the different video quality methods.We present the result here from six different well-known methods.A few commercially available methods were also considered, but since they did not perform better than the best performing methods in this investigation, we decided not to present these results.
Most evaluations for video quality uses naïve or non-expert viewers in their investigations.In ITU Recommendation (ITU-T Rec P.910) [9] it is recommended that the viewer should not be involved in video quality evaluations as part of their work.In comparisons between expert and non-expert viewers it has been shown that the general trends are similar, but the experts are more sensitive and give lower votes than the non-expert viewers, as reported by Speranza, et al. (2010) [66].They also pointed out that this could pose a challenge for objective models, since they in general are based on non-expert viewer data.
The video experts used in this investigation considered themselves as real experts with mean 4.5 and a small standard deviation, i.e., 0.6.All of them with excellent vision which means that they had both the knowledge and the vision appropriate for the task.They also rated the distortions to be typical and to a larger extent in a quintessential range.
The task given was to rate the video quality in a single stimulus procedure [9] with a hidden reference, meaning that the test persons do not know when the reference is shown and it is rated in the same way as the other videos in the test.There are more sensitive methods when it comes to seeing and rating distortions as e.g., DSIS [8], but the experiment was designed to mimic the viewing situation by the referees as they will not have any references to compare with.Furthermore, the single-stimulus method allows for getting ratings on a larger number of different videos in the same time period of time as DSIS [10].
The results show that even for video experts it is not possible to observe any quality degradation for the highest bitrates of H.264 for 1080p, as shown in Fig. 6.According to the results, the quality degradations are apparent first below 20 Mbit/s.The Motion JPEG quality on the other hand drops very quickly with decreasing bitrate and even highest bitrate clearly lower quality than the reference.Similarly, for 1080i but here 20 Mbit/s was found to be significantly different as well as the 10 Mbit/s for H.264, see Fig. 9. Motion JPEG has the same behaviour as for 1080p.A simple deinterlacing was included as a degradation and this was found to have really low quality.
The 540i was a bit different as it tested different scaling algorithms along with different bitrates Fig. 12.The different scaling algorithms follow each other closely.Lanczos got consistently higher ratings per bitrates than bilinear and nearest neighbour, but it was only statistically significant for 20 Mbit/s and 10 Mbit/s and only between Lanczos and bilinear.
The estimation of the SOS parameter a for the different formats were 0.18 (1080i), 0.19 (1080p) and 0.22 (540i).The values were similar especially for 1080p and 1080i.They were also in the range expected based on the findings by Hossfeld, et al. (2011) [65].The a value for 540i is a bit higher than expected, but the degradation were here of a different type.Also, the smaller picture size may have affected the results.
The evaluation of the objective video quality models followed the standardized procedure outlined in ITU-T Rec.P.1401 [37] and what the Video Quality Experts Group (VQEG) has been following in their evaluations, see e.g.[20].The RMSE has been the primary metric for the evaluation and the metric for which statistical significance has been calculated.There is often a non-linear relationship between the scores from the test persons and the objective scores [37], which was compensated partly with a monotonic non-linear fit between the MOS and the objective scores.Here we used a 3 rd order polynomial function, but others can also be used e.g., a logistic function.Pearson and Spearman correlation were also calculated to give a richer description of the results.The statistical significance testing was based on the RMSE.
From a performance perspective, we can see that the VQM_VFD got the lowest RMSE and highest correlation score among all the models for both 1080p and 1080i, see Tables 10 and 11.It was statistically significantly better than all the other models for 1080p, but not significantly better than VMAF and VQM_General for 1080i, as shown in Tables 12 and 13.In Netflix's original evaluation VMAF and VQM_VFD have the same performance.They targeted H.264 also but with lower bitrates, with a mixtures of resolutions, with DSIS and naïve observers [3].The second group is VMAF and VQM_General which have statistically the same performance i.e., they are not statistically significantly different from each, although VMAF shows slightly lower RMSE and higher correlation than VQM_General.The third group is SSIM, PSNR and VIF that show significantly worse performance.
Although VQM_VFD had the best performance, for practical reasons, the wide spread usage, implementation and general familiarity, along with the good performance, the method finally adopted was VMAF [1].

Conclusions
The research questions that this article was addressing were: 1) What levels of video quality are considered good or better for expert video professionals especially targeting sports content?2) Which objective quality models that have been developed by using video quality data from naïve test persons and most likely a wider and lower quality range, can well predict the quality in a high-quality range as defined by expert professionals from the broadcast industry?

A user study of video quality
A user study was conducted, to answer research question 1, with the participation of 25 Swedish video experts who assessed perceived video quality on a five-point scale, encompassing categories ranging from Excellent to Bad.The video content selected for evaluation focused on football-related scenes with dynamic motion and moving subjects.Three video formats were included: 1080p, 1080i, and 540i.The degradation processes mostly involved MJPEG and H.264 encoding, with bitrates ranging from 80 Mbit/s to 10 Mbit/s.Additionally, 1080i involved a simple deinterlacing error condition, and 540i incorporated three different scaling schemes in addition to the H.264 encoding.
The results revealed that MJPEG experienced a rapid decline in quality, with a significant decrease observed at 80 Mbit/s compared to the uncompressed reference.For H.264, no significant quality difference was observed until the bitrate dropped to 10 Mbit/s for 1080p, and 20 Mbit/s for 1080i.In the case of 540i, some scaling methods showed a significant quality decrease at 20 Mbit/s.Regarding deinterlacing for 1080i, careful consideration is crucial, as a low-quality deinterlacing scheme led to poor quality scores.Among the scaling schemes, Lanczos performed the best, while bilinear yielded the worst results.
The SOS analysis confirmed that the study's video quality assessments fell within the expected range for such experiments.Both 1080i and 1080p exhibited similar results, with MOS values slightly skewed towards higher quality.Conversely, for 540i, the distribution was more symmetric.

Objective video quality model performance
To answer research question 2, six different video quality models were independently evaluated for their performance with 1080p and 1080i videos.VQM_VFD demonstrated the best performance for both formats, followed by VMAF and VQM General models.On the other hand, SSIM, PSNR, and VIF exhibited similar, lower performance compared to the evaluated video models.Notably, SSIM's performance was particularly low for 1080i, mainly due to the HRC with a low-quality deinterlacing method.Scatter plots indicated that PSNR and VIF also faced similar issues in their performance assessments.

Future work
Future work will include evaluation of new broadcast formats such as 4 K and 8 K as well as high dynamic range video.Different frame rates will also be valuable to investigate.New encoders are also developed and should be included in future studies.The impact of zooming was not covered by the current study, but it is important to study as zoom is used frequently for a closer view of situations.Is it possible to find a perception limit when the remaining quality after zooming is not good enough to be used for decisions?New objective methods are also developed especially using machine learning techniques.This will potentially also enable the usage of NR methods for live quality monitoring, but this is still not mature enough to be used in production [19].This area is developing fast and new ways of processing the videos may prove to be very valuable [67].

Limitations
The uniqueness of this study lies in its focus on video quality opinions from video experts rather than end-users or consumers of broadcast TV.As such, its limitation is also a distinctive feature.The content primarily centres around sports, particularly Football, but additional content was included to broaden the relevance of the findings.The study encompassed video formats of 1080p, 1080i, and 540i, thus offering insights into these specific formats.However, the direct applicability of the results to other video formats may be limited.
It is essential to note that the study's scope is restricted to assessing the performance of the objective models considered within this investigation, which is contingent on the video content and codec employed in this study.and input on test person candidates for the subjective experiment.The authors acknowledge Pär Johansson (RISE) for proofreading, valuable discussions, and general support in the project as well as Jade Lumley for the thorough native language review.This work was funded by Fédération Internationale de Football Association (FIFA) and Sweden´s Innovation Agency (VINNOVA, dnr.2021-02107) through the Celtic-Next project IMMINENCE (C2020/2-2), which is hereby gratefully acknowledged.

Fig.
Fig. Stimulus presentation in the Absolute Category Rating (ACR) method (Figure borrowed from ITU-T P.910 [8]) Football_goal_2(used only in the training session)Overview over football goal area, involving several players there.Many people at a distance.Motion from camera pan and player's motion.

Fig. 15
Fig.15 Scatterplots of VMAF performance for 1080i.The left graph shows the results after the 3 rd order monotonic polynomial fit with a linear correlation line overlaid.The right graph shows the results before the 3 rd order monotonic polynomial fit with the fitted function overlaid

Fig. 16
Fig.16 Scatterplots of VQM General performance for 1080p.The left graph shows the results after the 3 rd order monotonic polynomial fit with a linear correlation line overlaid.The right graph shows the results before the 3 rd order monotonic polynomial fit with the fitted function overlaid

Fig. 18
Fig.18 Scatterplots of VQM_VFD performance for 1080p.The left graph shows the results after the 3 rd order monotonic polynomial fit with a linear correlation line overlaid.The right graph shows the results before the 3 rd order monotonic polynomial fit with the fitted function overlaid

Fig. 20 Fig. 21
Fig.20 Scatterplots of PSNR performance for 1080p.The left graph shows the results after the 3 rd order monotonic polynomial fit with a linear correlation line overlaid.The right graph shows the results before the 3 rd order monotonic polynomial fit with the fitted function overlaid

Fig. 22 Fig. 23
Fig.22 Scatterplots of SSIM performance for 1080p.The left graph shows the results after the 3 rd order monotonic polynomial fit with a linear correlation line overlaid.The right graph shows the results before the 3 rd order monotonic polynomial fit with the fitted function overlaid

Fig. 24 Fig. 25
Fig.24 Scatterplots of VIF performance for 1080p.The left graph shows the results after the 3 rd order monotonic polynomial fit with a linear correlation line overlaid.The right graph shows the results before the 3 rd order monotonic polynomial fit with the fitted function overlaid

Table 3
Summary of main pros and cons of previous studies.The pros and cons are here meaning suitability for our purpose and cons are not pointing out flaws in the investi-

Table 4
Description of the source video sequence of the 1080p formatThe camera follows the football in an almost empty stadium.Very few people.Motion mostly camera pan.

Table 5
Description of the source video sequence of the 1080i formatThe camera follows a player after he has scored a goal.One person up to a group of people.Closeups.Motion mostly camera pan.

Table 6
The HRCs or processing applied to the 1080p format

Table 7
The HRCs or processing applied to the 1080i format FFMPEG -c:v libx264 -profile:v high422 -b:v 10 M 10 Deinterlacing Uncompressed Blend fields, double the frame rate, top-field-first, VirtualDub (version 1.10.4)

Table 9
Hardware and software used

Table 10
Performances of the different models for 1080p.PCC (raw) is the Pearson Correlation Coefficient calculated on the raw scores.PCC (fit) is the linear correlation based on the 3 rd order monotonic polynomial fit.SCC is the Spearman rank-order correlation or Spearman's rho.RMSE is the Root Mean Square Error

Table 11
Performances of the different models for 1080i.PCC (raw) is the Pearson Correlation Coefficient calculated on the raw scores.PCC (fit) is the linear correlation based on the 3 rd order monotonic polynomial fit.SCC is the Spearman rank-order correlation or Spearman's rho.RMSE is the Root Mean Square Error

Table 12 P
[54]ues of statistical test on the difference in RMSE based on ITU-T Rec.P.1401[37]for 1080p.Significant values are marked with *, based on an alpha of 0.05 and the method of Holm[54]for multiple comparisons of15 comparisons

Table 13 P
[54]ues of statistical test on the difference in RMSE based on ITU-T Rec.P.1401[37]for 1080i.Significant values are marked with *, based on an alpha of 0.05 and the method of Holm[54]for multiple comparisons of 15 comparisons