Video coding standards normally define the syntax of bitstream and the decoding process, while encoders generate standard-compliant bitstream and thus determine compression performance. Each generation of video coding standard comes with a reference test model, such as HM (HEVC Test Model) for HEVC, which can be used to provide a performance benchmark. H.264/MPEG-4-AVC [ITU-T Rec H.264 (2005)] was launched in 2004, and is still the most prolific video coding standard, despite the fact that its successor, H.265/HEVC [ITU-T Rec H.265 (2015)] finalised in 2013, provides enhanced coding performance. In 2020, the first version of the latest video coding standard, Versatile Video Coding (VVC) [Bross et al. (2019)], has been finalised, which can achieve 30–50% coding gain over H.265/HEVC, supporting immersive formats (360° videos) and higher spatial resolutions, up to 16 K.

Alongside recent MPEG standardisation, there has been increasing activity in the development of open-source royalty-free video codecs, particularly by the Alliance for Open Media (AOMedia), a consortium of video-related companies. VP9 [Google (2017)], which was earlier developed by Google to compete with MPEG, provided a basis for AV1 (AOMedia Video 1) [AOM (2019); Chen et al. (2020)], which was released in 2018. Chen et al. (2020) set as AV1’s main goal to provide an open source and royalty-free video coding format that substantially outperforms its primary market competitors in compression efficiency and concurrently achieves a practical decoding complexity, is optimized for hardware feasibility and scalability on modern devices. Since its release, its pool of contributors has expanded and they have provided significant updates towards its goal. The AOM contributors are currently working towards developing the next generation, AV2 [Shostak et al. (2021)]. For further details on existing video coding standards and formats, the readers are referred to Bull and Zhang (2021); Wien (2015); Ohm (2015).

The performance of video coding algorithms is usually assessed by comparing their rate-distortion (RD) or rate-quality (RQ) performance on various test sequences. According to Recommendation ITU-R BT.500-12 (2012), the selection of test content is important and should provide a diverse and representative coverage of the video parameter space. Objective quality metrics or/and subjective opinion measurements are normally employed to assess compressed video quality. The overall RD or RQ performance difference between codecs can be then calculated using Bjøntegaard delta metrics for objective quality metrics according to Bjøntegaard (2001) or SCENIC developed by Hanhart and Ebrahimi (2014) for subjective assessments.

Most recent literature on video codec comparisons has focused on performance evaluations between MPEG codecs (H.264/AVC, HEVC, and VVC) and royalty-free (VP9 and AV1) codecs [Akyazi and Ebrahimi (2018); Grois et al. (2016); Lee et al. (2011); Dias et al. (2018); Nguyen and Marpe (2021)] and on their application in adaptive video steaming services Guo et al. (2018); Zabrovskiy et al. (2018); Katsavounidis and Guo (2018). However, the results presented are acknowledged to be highly inconsistent [Nguyen and Marpe (2021)], mainly due to the different configurations employed across codecs.

Our work contributes towards a fair codec comparison by releasing publicly a dataset that includes objective and subjective codec comparisons in three different use cases: encoding at FHD and UHD resolution and encoding within the framework of adaptive streaming.

The literature is rich of video datasets developed for many different purposes, mainly for computer vision related tasks such as object detection, action recognition, summarization, etc. These datasets, however, are not suitable for research in video compression. The main reason is that those have been designed for training deep network or models to infer very specific information. For example, the EPIC-Kitchens dataset [Damen et al. (2021)], contains scenes of daily activities in the kitchen, e.g., slicing bread, peeling carrots, stirring soup, etc. Therefore, the content bears significant similarities by repeating specific patterns in similar although diverse set-ups. Not providing a wide range of scenes with spatial and temporal information, this type of datasets cannot adequately represent video content and, thus, form a basis for a fair comparison of video compression algorithms and/or video quality metrics. Furthermore, the datasets published for computer vision tasks include already encoded versions (usually H.264-based encodings) as exported automatically from the video recording device. Although this type of content resembles the features of user generated content (UGC), a large portion of the videos streamed are coming from the creative industry. Thus research on video technology additionally requires pristine videos as exported from post-production.

In Table 1, a selection of commonly use datasets for video compression research is listed along with some of their basic features. Some of these datasets consist of pristine sequences [those that include raw (uncompressed) video sequences in the dataset] and others of UGC content [no raw sequences available as in KonViD-1K (Hosu et al. (2017)], YouTube-UGC [Wang et al. (2019)], DVL 2021 [Xing et al. (2022)]. Most of the existing datasets with pristine content, only include encoded sequences with one codec, usually H.264 {e.g., VQEGHD3 [Video Quality Experts Group (2010)], LIVE [Seshadrinathan et al. (2010)], NFLX-P and VMAF+ [Li et al. (2016)]} or HEVC {e.g., BVI-HD [Zhang et al. (2018)], BVI-Texture [Papadopoulos et al. (2015)], BVI-SynTex [Katsenou A. V. et al. (2021)]}.

From Table 1, it is evident that there is no dataset available that offers variations of encoded sequences based on different state-of-the-art codecs. Furthermore, to the best of our knowledge, there is no other dataset available that offers encoded sequences based on VVC and AV1. This is a very important contribution of the introduced dataset, BVI-CC, as it could facilitate research on video compression and comparison across these three different codecs. It is also noticeable that although almost all datasets provide opinion scores (OS) data, most of the datasets do not provide computed values of objective quality metrics (QMs).

Nine source sequences were selected from Harmonic (2019), BVI-Texture [Papadopoulos et al. (2015)] and JVET (Joint Video Exploration Team) CTC (Common Test Conditions) datasets. Each sequence is progressively scanned, at UHD resolution, with a frame rate of 60 frames per second (fps) and without scene-cuts. All were truncated from their original lengths to 5 seconds [rather than the recommended 10 s in ITU standard Recommendation ITU-R BT.500-12 (2012)]. This reflects the recommendations of a recent study on optimal video duration for subjective quality assessment by Mercer Moss et al. (2016a, 2016b). Sample frames from the selected nine source sequences alongside clip names and indices are shown in Figure 1. The dataset includes three sequences with only local motion (without any camera motion, V1-V3), three sequences with dynamic textures [for definitions see Zhang and Bull (2011), V4-V6], and three with complex camera movements (V7-V9). The coverage of the video parameter space is confirmed in Figure 2, where the Spatial and Temporal Information (SI and TI, respectively), the colourfulness (CF), and average contrast, as defined by Winkler (2012) are plotted.

In order to investigate coding performance across different resolutions and within an adaptive streaming framework, three spatial resolution groups were generated from the source sequences: (A) UHD (3840×2160) only, (B) HD (1920 601,080) only, and (C) HD-Dynamic Optimizer (HD-DO). For group C, coding results for three different resolutions (1920×1080, 1,280×720, and 960×540) and with various quantisation parameters (QPs) were firstly generated. The reconstructed videos were then up-sampled to HD resolution (in order to provide a basis for comparison with the original HD sequences). Here, spatial resolution re-sampling was implemented using Lanczos-3 filters designed by Duchon (1979). The rate points with optimal rate-quality performance based on Video Multimethod Assessment Fusion (VMAF) [Li et al. (2016)] were selected across the three tested resolutions for each target bit rate and codec. This process is repeated to create the entire convex hull in the DO approach [Katsavounidis (2018)]. The resulting selected resolutions across the set of target bit rates are reported in Table 2.

The reference test models of HEVC and VVC, and their major competitor, AV1 have been evaluated in this study. Each codec was configured using the coding parameters defined in their common test conditions [see Sharman and Suehring (2018); Bossen et al. (2019); Daede et al. (2019)], with fixed quantisation parameters (rate control disabled), the same structural delay (e.g., defined as GOP size in the HEVC HM software) of 16 frames and the same random access intervals (e.g., defined as IntraPeriod in the HEVC HM software) of 64 frames. The actual codec versions and configuration parameters are provided in Table 3.

Different target bit rates were pre-determined for each test sequence and for each resolution group (four points for resolution group A and B, and five for HD-DO group), and their values are shown in Table 4. These were determined based on the preliminary encoding results of the test sequences for each resolution group using AV1. This decision was made because the version of AV1 employed restricted production of bitstreams at pre-defined bit rates, as only integer quantisation parameters could be used. On the other hand, for HEVC HM and VVC VTM this was easier to achieve by enabling the “QPIncrementFrame” parameter. In order to achieve these target bit rates, the quantisation parameter values were iteratively adjusted to ensure the output bit rates were sufficiently close to the targets (within a range of ±3%).

In summary, a total number of 306 distorted sequences were produced: there are 108 (9 source sequences × 4 rate points × 3 codecs) for Resolution Group A (UHD only), 108 (9 3 4 co3) for Resolution Group B (HD only), and 90 (9 ) for 2) for Resolution Group C (HD-DO) ^{2, 3} .

All three experiment sessions were conducted in a darkened, living room-style environment. The background luminance level was set to 15% of the peak luminance of the monitor used (62.5 lux) Recommendation ITU-R BT.500-12 (2012). All test sequences were shown at their native spatial resolution and frame rates, on a consumer display, a SONY KD65Z9D LCD TV, which measures 1,429 29804 mm, with a peak luminance of 410 lux. The viewing distance was set to 121 cm (1.5 times the screen height) for Resolution Group A (UHD) and 241 cm (three times the screen height) for Resolution Group B (HD) and C (HD-DO), following the Recommendation ITU-R BT.500-12 (2012) and P.910 (1999). The presentation of video sequences was controlled by a Windows PC running an open source software, BVI-SVQA, developed by Dimitrov et al. (2019) at the University of Bristol for psychophysical experiments.

In all three experiments, the Double Stimulus Continuous Quality Scale (DSCQS) [Recommendation ITU-R BT.500-12 (2012)] methodology was used. In each trial, participants were shown a pair of sequences twice, including original and encoded versions. The presentation order was randomised in each trial and was unknown to each participant. Participants had unlimited time to respond to the following question (presented on the video monitor): “Please rate the quality (0–100) of the first/second video. Excellent–90, Good–70, Fair–50, Poor–30 and Bad–10”. Participants then used a mouse to scroll through the vertical scale and score (0–100) for these two videos. The total duration of each experimental session was approximately 50 (Resolution Group A and B) or 60 (Resolution Group C) minutes, and each was split into two sub-sessions with a 10 min break in between. Before the formal test, a training session was conducted, under the supervision of the instructor, consisting of three trials (different from those used in the formal test) to allow the participants time to familiarize.

A total of 62 subjects, with an average age of 27 (age range 20–45), from the University of Bristol (students and staff), were compensated for their participation in the experiments. All of them were tested for normal or corrected-to normal vision. Consent forms were signed by each participant and the data were anonymized. Responses from the subjects were first recorded as quality scores in the range 0–100, as explained earlier. Difference scores were then calculated for each tested sequence and each subject i by subtracting the quality score of the distorted sequence OS _dis from its corresponding reference OS _ref . Difference Mean Opinion Scores (DMOS) and the respective statistics (standard error, confidence intervals, etc) were then obtained for each trial by taking the mean of the difference scores among participants N. Particularly, DMOS values for a video v were calculated as follows: D M O S v = 1 N ∑ i = 1 N O S ref , i − O S dis , i . (1)

In order to reflect the quality taking into account the relative difference from the score of the hidden reference (in this case the original) for each participant i, we subtracted the DMOS _v from the maximum quality score (100).

Further to the score collection, we performed post-screening of the subjects. Following the Recommendation ITU-R BT.500-12 (2012) protocols, for each resolution group, we performed outlier rejection on all participant scores. No participants were rejected. Moreover, we calculated the subject bias according the Recommendation P.913 (2021) and removed it before performing any statistical analysis. As defined in the recommendation, subject bias is the difference between the average of one subject’s ratings and the average of all subjects’ ratings for each processed video sequence. To remove subject bias, the recommendation proposed to subtract that its value from each one of subject’s ratings.

Table 5 summarises the Bjøntegaard Delta measurements (BD-rate) of AOM AV1 (for three resolution groups) and VVC VTM (for Resolution Group A and B only) compared with HEVC HM, based on both PSNR and VMAF. For the tested codec versions and configurations, it can be observed that AV1 achieves an average bit rate saving of 7.3% against HEVC HM for the UHD test content assessed by PSNR, and this figure reduces (3.8%) at HD resolution. When VMAF is employed for quality assessment, the coding gains of AV1 over HM are slightly higher, averaging 8.6 and 5.0% for UHD and HD respectively. Comparing to AV1, VTM provides significant bit rate savings for both HD and UHD test content, with average BD-rate values between −27% and −30% for PSNR and VMAF. For resolution group C, where VMAF-based DO was applied for HM and AV1, the coding gain achieved by AV1 is 6.3% (over HM) assessed by VMAF, while there is a BD-rate (1.8%) loss when PSNR is employed. In overall conclusion, the performance of AV1 makes a small improvement over HM on the test content, and both AV1 and HM perform (significantly) worse than VTM.

In order to further compare performance across different spatial resolutions and within the context of DO, the average rate-VMAF curves of the nine test sequences (HD resolution only) for HM and AV1 with and without DO are shown in Figure 3. It can be observed that DO has achieved slightly higher overall coding gains for AV1 (BD-rate is -8.2%) on the tested content compared to HM (BD-rate is -6.1%). For both codecs, the savings become lower for higher bit rates (low QP and high quality). It should be noted that the DO approach employed was based on efficient up-sampling using simple spatial filters. Afonso et al. (2019), Zhang et al. (2019) and others have reported significant improvement when advanced up-sampling approaches are applied, such as deep learning based super-resolution.

As outlined in the introduction of Section 5, following the ITU-BT.500 protocols, for each resolution group, we performed outlier rejection on all participant scores. No participants were rejected. Next, following the ITU-T P.913 recommendation P.913 (2021), we estimated the subject bias for the different trials and compensated the raw scores based on that. Figures 4, 5 illustrate the 100-DMOS against the achieved bit rate for all tested codecs and sequences. Then, we performed one-way ANOVA analysis between pairs of the tested codecs to assess the significance of the differences. Table 6 summarises this comparison.

Results on Resolution Group A: A first impression from Figure 4 is that in most cases the VTM curve is on top of the other curves but also that the confidence intervals are overlapping in most cases. This is confirmed in Table 6, where the significance tests indicate only four cases which exhibit significant difference (p < 0.05) between HM and AV1. In two of these, AV1 is significantly better than HM: in the cases of CalmingWater sequence at R1 and LampLeaves at R2. Both of these sequences consist of dynamic textures that are very challenging to compress. The opposite happens at R2 for AirAcrobatics and at R1 for Myanmar sequence; two sequences with a static background and slow moving objects. Performing significance tests between VTM and HM, a higher number of cases with significant differences were identified mostly at the lower bit rates. Particularly, for CatRobot at R1, R3; for CalmingWater at R1; for LampLeaves at R1-R2; for DaylightRoad at R1-R3; and for RedRock at R1-R2. A similar number of cases where VTM was significantly better than AV1 were identified: for AirAcrobatics at R3; for CatRobot at R1; for Myanmar at R1; for ToddlerFontain at R1; for DaylightRoad at R1-R3; and for RedRock at R1-R2. It is worth mentioning that from the curves in Figure 4 VTM achieves very good quality ( ≥ 80 ) even at the lowest bit rates for all sequences except for the ones with dynamic textures (CalmingWater, ToddlerFontain, and LampLeaves). That remains an area for improvement for the future codecs.

Results on Resolution Group B: We performed the same post-screening of the subjects and statistical analysis as in resolution group A. Figure 5 demonstrates the subjective quality against the bit rate after removing the subject bias and Table 6 summarises the results of the one-way ANOVA comparison of the codecs. Generally, the results for the resolution group B align with those from resolution group A, that in most cases AV1 and HM result in equivalent video quality according to the viewers and that VTM in many cases prevails both codecs. Particularly, from Table 6, it can be observed that HM is significantly better than AV1 in only three cases: for Myanmar at R1-R2 and for RedRock at R2. VTM is significantly better than HM in 15 cases: for ToddlerFontain at R1-R4; for Myanmar at R1-R2; for CalmingWater at R1; for LampLeaves at R1-R2; for DaylightRoad at R1-R3; at RedRock at R1; and RollerCoaster at R3-R4. Similarly, VTM is significantly better than AV1 in 17 cases (almost have of the test sequences): for AirAcrobatics at R1; for CatRobot at R1-R3; for Myanmar at R1-R4; for LampLeaves at R1; for DaylightRoad at R1-R3; for RedRock at R1-R3; and RollerCoaster at R1-R2.

The reason that significant differences are noticed between AV1 and both HM and VTM in the case of the Myanmar sequence might be associated with the observation that, at lower bit rates, AV1 encoder demonstrates noticeable artifacts on regions of interest, namely in the center of the frame and on the heads of the walking monks in front of a still background with a static camera. An indicative example of these artifacts has been captured in Figure 6. This, however, is probably a rare case as no similar cases have been reported so far in recent literature.

Results on Resolution Group C: The removal of the subject bias leads to slightly different results than the ones presented for this case in our previous work [see Katsenou et al. (2019b)], as illustrated in Figure 7. After performing the significance test using one-way ANOVA between paired AV1 and HM sequences, the p-values of five rate points were indicated as significantly different. In all cases, HM is significantly better than AV1: at R1-R4 for Myanmar; and at R2 and R5 for LampLeaves. Myanmar encoded sequences suffer as mentioned earlier from unique artifacts and LampLeaves is a challenging dynamic texture. Although the findings from this resolution group are generally aligned with the observation from the other two resolution groups, for the LampLeaves sequence we notice a degraded performance of AV1 compared to resolution group B. This is attributed to the selected set of resolutions by the DO algorithm for AV1, which comprises lower resolution sequences than those selected for HM: at R2 540p instead of 720p and at R5 720p instead of 1080p (see Table 2).

The correlation performance of six tested objective quality metrics for three resolution groups (in terms of SROCC values) is summarised in Table 7. It can be observed that VMAF outperforms the other metrics on all three test databases with the highest SROCC and LCC values, and lowest OR and RMSE. PSNR results in much lower performance, especially for the UHD resolution group. It is also noted that, for all test quality metrics, the SROCC values for three resolution groups are all below 0.9, which indicates that further enhancement is still needed to achieve more accurate prediction.

The average complexity figures for encoding UHD and HD content are summarised in Table 8, where the HM encoder has been used for benchmarking. The average complexity is computed as the average ratio of the execution time of the tested codec for all rate points over the benchmark. As can be seen, for the tested codec versions, AV1 has a higher complexity compared to VTM ⁴ . Interestingly these figures are higher for the HD than the UHD resolution. The relationship between the relative complexity and encoding performance (in terms of average coding gains for PSNR and VMAF) is also shown in Figure 8.