Construction and Evaluation of a Video Retrieval System Using Animal Recognition and Conceptual Relationship between Family and Species

Abstract

In recent years, video streaming services have gained widespread popularity. Typically, the search function on such platforms assesses the relevance of a query based on the video’s title, tags, description, and other metadata provided by the creator. As a result, if a video’s title does not accurately reflect its content, users may encounter irrelevant search results. To address this issue, we developed a novel system that retrieves animal videos based on their content using image recognition technology. To allow users to intuitively search for ambiguous information, we incorporated a hierarchical relationship between animal families and species, enabling searches by family name. We also developed and tested an application that facilitates efficient video retrieval, allowing users to view search results as a digest video and navigate from the digest to the original video. In an evaluation experiment focusing on the user’s prior knowledge, the system achieved a task completion rate of 0.95 when the user knew the family name but not the species name, indicating that the system is effective in retrieving relevant scenes even when the user has imprecise information.