Topic-Oriented Text Features Can Match Visual Deep Models of Video Memorability

Not every visual media production is equally retained in memory. Recent studies have shown that the elements of an image, as well their mutual semantic dependencies, provide a strong clue to whether video clip will be recalled on second viewing or not. We believe short textual descriptions encapsulate most these relationships among video, and thus they represent rich yet concise source information tackle problem memorability prediction. In this paper, we deepen study captions means convey natural language semantics video. propose use vector embeddings from pretrained SBERT topic detection model with no adaptation input features linear regression model, showing that, such representation, simpler algorithms can outperform deep models. Our results suggest text expressed might effective embodying required memorability.