Video quality assessment for adaptive streaming over TCP


The largest amount of Internet bandwidth is today used for video streaming, in particular from YouTube and Netflix. These use a technology call adaptive streaming over TCP, which tries to adapt as smoothly as possible to the available network capacity and glitches such as packet loss. Researchers strive to keep the video quality stable for the longest possible period in time, only reducing it to avoid that the video playback has to stop. However, the research showing that this is actually necessary for the best possible user experience has just started. This thesis is meant to contribute to it.

Description of topic

In the networking literature, video quality is most evaluated using a quality metric that doesn't take temporal effects into account at all. Mostly, this means that papers are written that use the peak signal-to-noise ration (PSNR) to argue that one video leads to more user satisfaction than another. Interestingly, it has first been shown in 1974 that this does not even work for images "The effects of a visual fidelity criterion of the encoding of images". A series of standards exist that propose metrics that have been proven to be much better suited to evaluate video quality, recently the standard ITU.T J.341.

But now, with adaptive streaming over TCP, a question arises that has not been asked yet. It is:
"If video quality changes rarely but in fairly big jumps, how strong is the penalty for changing quality?"
A paper addressing this question has been written already, but this is just the start of a possibly long exploration: "Subjective Quality Assessment of an Adaptive Video Streaming Model". 


This thesis is meant to contribute to this exploration. The thesis will proceeds by exploring two main points:
1. investigate through literature study how to investigate user's opinion in ranges of minutes
2. conduct a study that tests how users' opinion about video quality changes on the range of minutes


  • Pål Halvorsen
  • Carsten Griwodz