Video Area Home >> Video Codecs Comparisons >> MSU Video Codecs Comparison 2021

MSU Video Codecs Comparison 2021 Part 2: Subjective

Sixteen Annual Video Codecs Comparison by MSU

Video group head: Dr. Dmitriy Vatolin
Project head: Dr. Dmitriy Kulikov
Measurements, analysis: Dr. Mikhail Erofeev,
Anastasia Antsiferova,
Egor Sklyarov,
Alexander Yakovenko,
Nickolay Safonov,
Alexander Gushin,
Nikita Alutis
compression.ru Lomonosov
Moscow State University (MSU)
Graphics and Media Lab
Dubna International
State University
Institute for Information
Transmission Problems,
Russian Academy of Science

News

  • 23.11.2021 Report release

Navigation


Results


  • The places below are given only for quality scores, not taking encoding speed into account
Slow (1 fps) Fast (30 fps)
Best quality
(YUV-Subjective)
1st: S266
2nd: Aurora AV1
3rd: QAVS3
1st: Tencent V265
2nd: QAV1
3rd: Phoenix265


For subjective quality measurements we used Subjectify.us crowdsourcing platform. We involved more than 10,800+ participants

Download and buy report


Free Enterprise
Number of test sequences 3 of 15 (only Universal use case and only 8-bit content) All 15 sequences for all use cases
Test video descriptions
Basic codec info
Metric: YUV-Subjective, YUV-SSIM, VMAF (overall results)
Other objective metrics (Y-VMAF(0.6.1 for 4K), Y-SSIM, U-SSIM, V-SSIM, YUV-PSNR, Y-PSNR, U-PSNR, V-PSNR)
Test videos download
Encoders presets description
PDF report 67 pages 80 pages
HTML report 30 interactive charts 3000+ interacive charts
Price Free 950 USD
Download/Buy
PDF & HTML reports (ZIP)
You will receive enterprise versions of all 2021 reports (FullHD, Subjective, 4K)

Participated codecs


Codec name Use cases Standard Version
1 Aurora AV1
Visionular
Slow (1 fps) AV1 Linux
2 Phoenix265
Open Visual Cloud
Slow (1 fps),
Fast (30 fps)
H.265/HEVC v4.4, Linux
3 QAV1
iQIYI Inc.
Slow (1 fps),
Fast (30 fps)
AV1 v2.1, Linux
4 QAVS3
iQIYI Inc.
Slow (1 fps) AVS3 v1.0.0, Linux
5 rav1e
The rav1e contributors
Slow (1 fps) AV1 0.5.0-alpha (p20210518), Windows
6 Reference x265
MulticoreWare, Inc.
Slow (1 fps),
Fast (30 fps)
H.265/HEVC 3.5+1-f0c1022b6, Windows
7 S266
Alibaba Group
Slow (1 fps) H.266/VVC v1.1, Windows
8 SIF Codec
SIF Codec LLC
Slow (1 fps) SIF v1.0, Windows
9 SVT-HEVC
Open Visual Cloud
Slow (1 fps),
Fast (30 fps)
H.265/HEVC 1.5.1, Windows
10 SVT-VP9
Open Visual Cloud
Slow (1 fps),
Fast (30 fps)
VP9 0.3.0, Windows
11 Tencent V265
Tencent
Slow (1 fps),
Fast (30 fps)
H.265/HEVC 1.5.0, Linux
12 VVenC
Fraunhofer HHI
Slow (1 fps) H.266/VVC v1.0.0, Linux
13 x264
x264 project
Slow (1 fps),
Fast (30 fps)
H.264/AVC r3065-ae03d92, Windows
14 x265
MulticoreWare, Inc.
Slow (1 fps),
Fast (30 fps)
H.265/HEVC 3.5+1-ce882936d, Windows
15 xin26x (HEVC)
A Father (xin26x)
Slow (1 fps),
Fast (30 fps)
H.265/HEVC v1.0, Windows
16 xin26x (VVC)
A Father (xin26x)
Slow (1 fps),
Fast (30 fps)
VVC v1.0, Windows

Subjective Comparison Methodology


For subjective quality measurements we used Subjectify.us crowdsourcing platform. We involved 10,800+ participants. After deleting replies from bots we got 529,171 pairwise answers. Bradley-Terry model was used to compute global rank.

To conduct an online crowdsourced comparison, we uploaded encoded streams to Subjectify.us. For better browser compatibility we performed transcoding with x264 and CRF=16.

The platform hired study participants and showed the upload streams to them in pairs. Each pair consisted of two variants of the same test video sequence encoded by various codecs at various bitrates. Videos from each pair were presented to study participant sequentially (i.e., one after another) in full-screen mode. After viewing each pair, participants were asked to choose the video with the best visual quality. They also had the option to play the videos again or to indicate that the videos have equal visual quality. We assigned each study participant 12 pairs, including 2 hidden quality-control pairs, and each received money reward after successfully completing the task. The quality-control pairs consisted of test videos compressed by the x264 encoder at 1 Mbps and 4 Mbps. Responses from participants who failed to choose the 4 Mbps sequence for one or more quality-control questions were excluded from further consideration.

In total we collected 529,171 valid answers from 10,800+ unique participants. To convert the collected pairwise results to subjective scores, we used the Bradley-Terry model [1]. Thus, each codec run received a quality score. We then linearly interpolated these scores to get continuous rate-distortion (RD) curves, which show the relationship between the real bitrate (i.e., the actual bitrate of the encoded stream) and the quality score. Section "RD Curves" shows these curves.

We obtained the subjective scores for this study using Subjectify.us. This platform enables researchers and developers to conduct subjective comparisons of image and video processing methods (e.g., compression, inpainting, denoising, matting, etc.) and carry out studies of human quality perception.

To conduct a study, researchers must apply the methods under comparison to a set of test videos (images), upload the results to Subjectify.us and write a task description for study participants. Subjectify.us handles all the laborious steps of a crowdsourced study: it recruits participants, presents uploaded content in a pairwise fashion, filters out responses from participants who cheat or are careless, analyzes collected results, and generates a study report with interactive plots. Thanks to the pairwise presentation, researchers need not invent a quality scale, as study participants just select the best option of the two.

The platform is optimized for comparison of large video files: it prefetches all videos assigned to a study participant and loads them into his or her device before asking the first question. Thus, even participants with a slow Internet connection won’t experience buffering events that might affect quality perception.
To try the platform in your research project, reach out to www.subjectify.us. This demo video shows an overview of the Subjectify.us workflow.


Codec Analysis and Tuning for Codec Developers and Codec Users


Computer Graphics and Multimedia Laboratory of Moscow State University:

  • 17+ years working in the area of video codec analysis and tuning using objective quality metrics and subjective comparisons.
  • 30+ reports of video codec comparisons and analysis (H.265, H.264, AV1, VP9, MPEG-4, MPEG-2, decoders' error recovery).
  • Methods and algorithms for codec comparison and analysis development, separate codec's features and codec's options analysis.

Strong and Weak Points of Your Codec

  • Deep encoder parts analysis (ME, RC on GOP, mode decision, etc).
  • Weak and strong points for your encoder and complete information about encoding quality on different content types.
  • Encoding Quality improvement by the pre and post filtering (including technologies licensing).

Independent Codec Estimation Comparing to Other Codecs for Different Use-cases

  • Comparative analysis of your encoder and other encoders.
  • We have direct contact with many codec developers.
  • You will know place of your encoder between other newest well-known encoders (compare encoding quality, speed, bitrate handling, etc.).

Encoder Features Implementation Optimality Analysis

We perform encoder features effectiveness (speed/quality trade-off) analysis that could lead up to 30% increase in the speed/quality characteristics of your codec. We can help you to tune your codec and find best encoding parameters.

Thanks


Special thanks to the following contributors of our previous comparisons
Apple Google Intel NVidia
Huawei AMD Adobe Tencent
Zoom video communications Facebook Inc. Netflix Alibaba
KDDI R&D labs Dolby Tata Elxsi Octasic
Qualcomm Voceweb Elgato Telecast
ATI MainConcept Vitec dicas

Contact Information

We appreciate any feedback on our comparison


Subscribe to report updates




Other Materials


Video resources:

Last updated: 12-May-2022


Server size: 8069 files, 1215Mb (Server statistics)

Project updated by
Server Team and MSU Video Group

Project sponsored by YUVsoft Corp.

Project supported by MSU Graphics & Media Lab