Title:Fast Automatic Video Retrieval using Web Images

Abstract:We describe a real-time video retrieval framework based on short text input in which weakly labelled training samples from the web are obtained, after the query is known. Concept discovery methods in such a setting train hundreds of detectors at test time and apply them to every frame in the video database. Hence, they are not practical for use in a text based video retrieval setting. We show that an efficient visual representation for a new query can be constructed on-line that enables matching against the test set in real-time. We evaluate a few combinations of encoding, pooling, and matching schemes that are efficient and find that such a system can be built with surprisingly simple and well-known components. We are not only able to construct and apply query models in real-time, but with the help of a re-ranking scheme, we also outperform state-of-the-art methods by a significant margin.