Abstract:
The thesis discusses and evaluates a model of video information retrieval
that incorporates a variation of Relevance Feedback and facilitates
object-based interaction and ranking. Object-based feature search
for video IR is one of the main novel aspects of this work.
Video and image retrieval systems suffer from poor retrieval performance
compared to text based information retrieval systems and this is mainly
due to the poor discrimination power of visual features that provide the
search index. Relevance feedback for video retrieval can help overcome
the poor discrimination power of the features with user essentially
pointing the system in the right direction based on their judgements.
The ostensive relevance feedback approach explored in this work weights
user judgements based on the order in which they are made with newer
judgements weighted higher than older judgements.
A user experiment has been developed in which three video retrieval
system variants are evaluated on a corpus of video content. All three
systems offer the following features for query search:
- character objects
- closed caption text
- colour and edge histogram
- user sketch
the second and third apply ostensive relevance feedback with
variations in order weight. In order to evaluate effective object
retrieval animated video content provides the corpus content for
the evaluation experiment.
The main findings in this thesis:
- Exploring the benefit of ostensive relevance feedback for video retrieval the work demonstrated that ostensive RF provided a performance boost for narrow query topics where few valid results are contained in the evaluation corpus.
- Examining the effectiveness of object retrieval the main finding was that the usage and performance of the object feature was high but tended to be topic specific.
- Users interaction approach was also explored. The main results demonstrate that users RF iterations tend to be small (typically 8 or less), text was the most popular search feature and objects were used in over 50% of query searches
To read the full Thesis, contact me at
pbrowne@computing.dcu.ie
