David Scott - PhD Transfer Talk - 1st June 2012

Video Category: 
Transfer Talk
DSC00377-half.JPG

Title: Enhancing Visual Representations in Multi-Modal Video Information Retrieval for Known-Item Search

Abstract: In recent years we have witnessed a surge in the amount of online content access on various handheld devices, this is largely attributed to the smart phone revolution and is set to continue with increasing numbers favouring the on-the-go nature of these always on devices. There comes new challenges with developing video retrieval systems associated with these devices and in respect to their visual representation is where the remit of my PhD lies. My proposed research focuses on the representation of content, specifically that of user-generated video content on handheld Information Retrieval system. Building on previous work for the TRECVid conference in both 2010 and 2011 where we developed a retrieval system on an iPad, we utilize the lessons learned in both use case and representation to employ three methods of clustering to determine how best to display the relevant keyframes to a user without the need for technical knowledge of content based retrieval systems.

Our three proposed system to be evaluated will comprise of 1. Best First Match - utilize only the first occurrence of multi keyframe representation where it lies in the clustered list regardless of rank. 2. Ignore Low Ranks - Low ranking keyframes from either text indexes or from the concepts will be ignored to speed up the search and to remove results which the system has deemed unnecessary. Also each video will only have one keyframe per cluster hit. 3. Catch All - the final approach is to use all keyframes in clustered ranked order.' this will allow for the reduction in user overhead for search and interaction by quantifying the best shots by using user-supplied evidence to determine ideal keyframe representation. Furthermore with the aid of clustering techniques I reduce user searching/browsing overhead by grouping and optimally representing similar content.