The number of bandwidth-hungry applications and services is constantly growing. HTTP adaptive streaming of audio- visual content accounts for the majority of today’s internet traffic. Although the internet bandwidth increases also constantly, audio-visual compression technology is inevitable and we are currently facing the challenge to be confronted with multiple video codecs. This paper provides a practical evaluation of state of the art video codecs (i.e., AV1, AVC/libx264, HEVC/libx265, VP9/libvpx-vp9) for large- scale HTTP adaptive streaming services. In anticipation of the results, AV1 shows promising performance compared to established video codecs. Additionally, AV1 is intended to be royalty free making it worthwhile to be considered for large scale HTTP adaptive streaming services.

1. A Practical Evaluation of Video Codecs for
Large-Scale HTTP Adaptive Streaming Services
2018 IEEE International Conference on Image Processing

2. Video codecs
- AVC/libx264
- HEVC/libx265
- VP9/libvpx-vp9
- AV1
Currently, we are in a situation where we can
choose from multiple video codecs such as
AVC, HEVC, VP9 and AV1.
Every new major video codec has increased
coding efficiency significantly compared to its
predecessor.
https://digitaltvnewssummary.files.wordpr
ess.com/2017/08/ah_1c-1.png?w=723

3. Research goal
Present coding performance evaluation of AV1 compared to AVC,
HEVC, and VP9 in the context of HTTP adaptive streaming
Create Multi-codec DASH dataset

4. The test video sequences

5. The average spatial information (SI) and
temporal information (TI) for video sequences

6. The bitrate ladder and segment length
Segment
2 sec.
Segment
4 sec.
t

7. Encoding and Packetization
AVC, HEVC and VP9
ffmpeg -y -i {input.y4m} -r 24 -vf scale={WxH} format=yuv420p -c:v
{libx264,libx265,libvpx-vp9} -preset slow -pass {1|2} -b {bitrate}
{out.mp4|out.webm}.
AV1
aomenc {input.y4m} -v --good --cpu-used=2 --width={W} --height={H} --
target-bitrate={bitrate} --psnr -o {out.webm}.
Packetization
- Two-pass encoding
- MP4 segments for AVC and HEVC
- WebM container format for VP9 and AV1

8. wPSNR and BD-rate
The weighted Peak Signal-to-Noise Ratio (wPSNR) for the luminance (Y)
and chrominance (UV) components
PSNR and bitrate are averaged over all frames in each sequence and the
Bjøntegaard-Delta bit-rate (BD-rate) is calculated from these values over
the entire bitrate ladder.

9. Y-PSNR results, 2 sec. segments
- AVC shows the lowest performance followed by VP9 although sometimes VP9 slightly outperforms HEVC
- AV1 shows in most cases better results compared to the other codecs, specifically for higher bitrates
- In some cases and specifically for lower resolutions HEVC is able to outperform AV1

10. Evaluation results
- AV1 codec is able to outperform all
the other codecs. The average
gains over AVC are around 48%,
17% compared to HEVC, and
13% compared to VP9.
- Gains vary strongly and depend on
the individual dataset sequences
- For HEVC, the gains range from
close to zero up to 31%
The BD-rate results for the segments of AV1 over the
entire bitrate ladder compared to other codecs

11. Evaluation results
For the high resolution (3840x2160)
and high bitrates, the coding gain of
AV1 compared to the other codecs is
even higher with an average of
roughly 58% compared to AVC,
44% compared to HEVC, and 27%
compared to VP9

12. Multi-codec dash dataset
Our dataset is available here: http://dash.itec.aau.at/
The goal of such a dataset is to enable interoperability testing and experimenting with
different adaptation strategies of DASH clients supporting multiple video codecs.
Anatoliy Zabrovskiy, Christian Feldmann, Christian Timmerer, Multi-codec DASH dataset, In MMSys '18
Proceedings of the 9th ACM Multimedia Systems Conference, ACM Press, New York (NY), pp. 438-443,
2018.

13. Thank you. Questions?
Anatoliy Zabrovskiy
anatoliy.zabrovskiy@aau.at