Shot Detection in Screencasts of Web Browsing with Convolutional Neural Networks[go to overview]
Daniel Vossen will defend his bachelor thesis about “Shot Detection in Screencasts of Web Browsing with Convolutional Neural Networks”. The talk is open for the university audience. Due to the current situation, everybody who wants to attend the talk must register via E-mail to email@example.com until 24th September, so we know who will attend and how many people to expect. See the official statement by university for information how to behave on campus in the current situation: https://www.uni-koblenz-landau.de/de/coronavirus
Behavior analysis of users on a graphical interface, which can change its appearance by interactions or background scripts, is used to evaluate the interface and to be a help in its future designing process. The challenge is to turn data of interactions with an interface into comprehensible and useful information. In this work, we take a closer look into screen recordings of interactions with websites and how visual changes in the interfaces can be detected. Detecting visual change is a key part of the aggregation of interaction data of several users onto a common space for later analysis. For this, we observe how visual change gets detected in traditional videos and whether the same methods can find a use for recordings of interaction with websites. We find that, while artificial neural networks tasked with the detection of visual change get used with great success for videos recorded with cameras, there exists no such work for recordings of interactions of websites yet. Therefore, in this bachelor thesis, we use a convolutional neural network to detect visual changes in recordings of interactions of websites and compare our results with other state-of- the-art procedures. We find out that in some cases we can reach comparable or even better results as state-of-the-art approaches but most of the time our approach does not offer a reasonable alternative.
25.09.20 - 12:15