Human action video retrieval
Páez Rivera, Fabián Mauricio · 2019
Abstract The problem of efficiently answering a user information need in a video collection related to human actions is addressed in this thesis. The focus is given to the case where the user queries are stated using an example video containing the action of interest. Among the motivations of the work is the growing complexity of available video content in terms of size and content diversity, and also the ubiquity of video content fueled by the widespread use of video cameras. To solve the problem at hand, an information retrieval system is proposed where multiple information modalities are leveraged if available to discover the latent semantics of the video collection. The central component are matrix factorization-based indexes which have been previously used on image retrieval settings. Along the way, different features and encoding methods for the visual information have been evaluated, such as Bag of Features, Fisher Vectors and Improved Trajectory Features. As a result, a system achieving similar performance as Support Vector Machines-based systems has been obtained.