Live transcoder that self-adjusts to process only the video files that will be requested by the viewers

Investigating the feasibility of using neural networks to reduce video processing overhead in HTTP-based live streaming is the objective of Jonathan Hedin’s Master Thesis project. A project that will be performed together with Eyevinn Technology in Stockholm as an R&D Sustainability project.

HTTP-based live streaming is the primary technology for distributing live streams over Internet today. A technology fundamentally based on open web standards which is one of the reasons this technology is the dominating one today. The way this technology works is that a live transcoder generates 2-3 seconds chunks of video files, where each video chunk is also generated in different video resolutions and file sizes. These video files are placed on a web server and the video player downloads one of these video files that it estimates it can download before the next one is produced. If the available bandwidth is low, it will choose one of the smaller video files available. While this approach in general works very well there is a processing overhead in terms of video files being produced that may never be downloaded or requested. Not only leads this to storage waste but also video processing / CPU waste.

Another, and relatively new, approach is to only generate the video files that are requested, by transcoding and creating these chunks on request. While this approach removes all waste it might add additional latency. Some video files will always need to be produced but it may vary depending on factors such as how many are viewing the stream, what bandwidth capability these viewers currently have etc. What if a live transcoder would generate in advance the video files most likely to be accessed and the rest are generated on request, and how would then the transcoder know in advance what video files are most likely to be accessed?

What this master thesis project will explore is whether player analytics data from the viewers can be used with machine learning to train a live transcoder what video files are most likely to be accessed and can be generated in advance to reduce the amount of video files needed to be generated on request without producing any waste.

“I personally believe it is both interesting and important to explore wherever you have the possibility to reduce overhead and energy footprint. And I also believe to early test and validate the feasibility of an idea. That is why I find this master thesis topic that Jonathan will explore really interesting, and it is perfectly in line with our sustainability area in R&D.”, says Jonas Birmé VP R&D Eyevinn Technology

Answering the question why he chose this topic Jonathan says:

"Video, image and sound are all interesting areas to me and video streaming undoubtedly have huge effects on society and the world considering the sheer scale of video content being produced and consumed each day. The ease of access to video, I think, although excellent from a usability perspective, can obscure the energy and resources needed to facilitate this. Therefore, finding ways to make video streaming more efficient seems like a worthwhile area to investigate. Eyevinn seemed like a good fit for this considering their experience in video streaming and focus on sustainability."

The collection of player analytics will be based on the open-source player analytics framework by Eyevinn and any added improvements made in this master thesis project will be contributed back to this open-source framework.

The master thesis project is planned to be completed by end of Q2 2023 if everything goes according to plan.

For more information about this press release contact:

Jonas Birmé
VP R&D Eyevinn Technology
E: jonas.birme@eyevinn.se