MediaFinder

The MediaFinder content-understanding suite consists of a set of unique deliverable solutions for a range of applications. It provides recommendations through a retrieval-by-example process that thinks like a human thinks, rather than for the ANN to tell you what it believes to be reality. MediaFinder was an award recipient from UK Innovate, the U.K.’s national innovation agency.

While most existing media content understanding frameworks focus on so called tagging or categorisation tasks (e.g., is this a cooking, beauty, animal or travel video?), EPAI’s unique approach acknowledges that human perception and perceived similarity between data records is a continuous space rather than a static set of predefined taxonomies. MediaFinder is able to make use of extracted physical properties to identify similarities in semantic content and so is independent of metadata.

Instead of assigning one or multiple classification tags to a given data record, our custom neural network extracts a so-called embedding, i.e., a unique fingerprint in the form of a series of vectoral representation or a data point. The key innovative property is that records which are perceived as similar by humans will have numerically similar embeddings, whereas records which are judged as being dissimilar by humans will result in embeddings which are numerically dissimilar. MediaFinder is both fast and scaleable and has unlimited granularity.

The resulting embedding space sets the basis for a broad range of applications, including recommendation engines, query-by-example or free-text search in media collections and the principle can be applied to different media types, including video, image and music.