How should Netflix and YouTube adapt video quality?


Netflix and YouTube are just two of many commercial and free services that make use of adaptive video streaming over HTTP (HAS), a generic technique that conquered the world of video-on-demand in just a few years. HAS is successful because it lets the browser (plugin) decide which quality to retrieve, and also because HAS makes it impossible to implement the really bad idea of adapting video quality to network conditions for every single frame.

But network conditions change over time, and adaptation is unavoidable. How to adapt, however, is a big question. A general consensus is that as few quality change events as possible are best, as long as video playback that stops because of empty buffers can be avoided. This big problem is that this general consensus has not been proven scientifically yet. The reason is that there is no accepted method for keeping track of people’s opinion as it changes while they watch videos for 45 to 90 minutes.

The method that is developed during this thesis must be feasible from a computer science point of view, it must be sound from the point of view of cognitive psychology, and it must be statistically correct.

We hope that the method can be applied in a crowdsourcing scenario, but if this turns out to be infeasible, it will be applied in a lab setting.


  • A new method for long-duration perceptual studies.
  • Results of subjective tests made available to the HAS research communities.

Learning outcome:

  • Understand how HAS protocols like Microsoft Smooth Streaming, Apple HLS, MPEG DASH, etc. work
  • Develop a test suite.
  • Understand how subjective quality assessment studies are conducted.
  • Understand how crowdsourcing tasks can be designed and conducted.


  • data communication
  • basic knowledge of math incl. statistics
  • basic knowledge of video coding



Contact person:

Carsten Griwodz

(Pål Halvorsen, Ragnhild Eg)