|Title||:||Spatio-Temporal Analysis of Video Shots for Content Based Video Retrieval (CBVR)|
|Speaker||:||Chiranjoy Chattopadhyay (IITM)|
|Details||:||Tue, 26 Aug, 2014 4:00 PM @ BSB 361|
|Abstract:||:||Content Based Video Retrieval (CBVR) is the problem of retrieving a set of videos from the database (gallery) which are similar to a given query video shot. Some of the application areas of CBVR are, video on-demand, event identification, near-duplicate detection, multimedia content analysis and video classification etc. Existing CBVR systems lack in efficiently computing a compact, unified (joint) representation of spatio-temporal information. Research issues involve: efficient tracking in moving camera videos, multiple moving objects, suitable weightage for trajectory and spatio-temporal information, matching criteria etc.
This presentation describes a few methods designed to overcome some of the key issues related to the CBVR task. Highlights of this talk are the following:
Segmenting the prominent moving foreground object in a video shot, captured using a moving camera, using iterative refinement over graph-cut. Novel video content descriptors (DMST-CSS, time series representation) for CBVR task. An approach towards solving the CBVR task for action videos and complex shots with multiple moving objects, using latent semantic indexing of spatio-temporal features. Results will be shown using real-world datasets, to exhibit the superiority of the proposed methods.