Abstract
Traditionally, the process of choosing the best thumbnail is having human curators or editors select it. It requires technical expertise as well as a good notion of which specific target population that content should be directed to. We present an algorithm that automatically extracts the “Best Frame” from the drama series’ episodes to be the thumbnail using not only the video but also the episode summary text and closed captions (CC). This “Best Frame” carries the essential characteristics that human professionals would look for while trying to manually select a thumbnail.
Introduction
Globoplay is an Over-the-top platform developed by GLOBO that offers access to broadcast content directly over the internet: streaming live content, drama series, documentaries, news and entertainment programs. Each video published in Globoplay is accompanied by a thumbnail and a short summary. In the case of drama thumbnails, to increase engagement, it is common to picture a relevant event that happened on that particular content. Globoplay’s interface can be seen as an example in Figure 1.
Like in other digital video platforms, it takes special care in the way products are presented to the users in order to increase their interest and engagement. In this scenario, thumbnails play a very important role in online video display. As the most representative snapshot, they are supposed to capture the essence of each video and provide an accurate first impression to the viewers. Studies suggest that people look at thumbnails extensively when browsing online videos. An accurate thumbnail would ultimately make a video more attractive to watch, which, in turn, leads to an increase in ad revenue.
This work introduces a novel algorithm responsible for the automatic thumbnail selection of drama series content published on Globoplay. The presented method uses not only image features but also text metadata to make the best thumbnail choice. It aims to achieve the same level of quality obtained by the manual process. In the next section, the problem context will be detailed further; following, the process of automation and its algorithm is explained; after that, some analysis results are shown; and finally, the work conclusions and future works are presented.
No comments yet