Efficient thumbnail identification through object recognition

Given the overwhelming growth of online videos, providing suitable video thumbnails is important not only to influence user’s browsing and searching experience, but also for companies involved in exploiting video sharing portals (YouTube, in our work) for their business activities (e.g., advertising). A main requirement for automated thumbnail generation frameworks is to be highly reliable and time-efficient, and, at the same time, economic in terms of computational efforts. As conventional methods often fail to produce satisfying results, video thumbnail generation is a challenging research topic. In this paper, we propose two novel approaches able to provide relevant thumbnails with the minimum effort in terms of time execution and computational resources. The proposals rely on an object recognition framework which captures the most topic-related frames of a video, and selects the thumbnails from its resulting frames set. Our approach is a trade-off between content-coverage and time-efficiency. We perform preliminary experiments aimed at assessing and validating our models, and we compare them with a baseline compliant to the state-of-the-art. The assessments confirm our expectations, and encourage the future improvement of the proposed algorithms, as our proposals are significantly faster and more accurate than the baseline.