Shot-based object retrieval from video with compressed Fisher vectors

This paper addresses the problem of retrieving those shots from a database of video sequences that match a query image. Existing architectures are mainly based on Bag of Words model, which consists in matching the query image with a high-level representation of local features extracted from the video database. Such architectures lack however the capability to scale up to very large databases. Recently, Fisher Vectors showed promising results in large scale image retrieval problems, but it is still not clear how they can be best exploited in video-related applications. In our work, we use compressed Fisher Vectors to represent the video-shots and we show that inherent correlation between video-frames can be proficiently exploited. Experiments show that our proposal enables better performance for lower computational requirements than similar architectures.