This thesis presents weakly supervised learning approaches to directly exploit image-level tags (e.g. objects, attributes) for comprehensive image understanding, including tasks such as object localisation, image ...