Leveraging from group classification for video concept detection

The performance of a content based retrieval system is limited mainly because of the unavailability of sufficient annotated examples, descriptor noise and the semantic gap that is the representation difference between the high level concept and the low level feature. Finding the optimal parameters of the learner for each concept adds to the difficulty of this task. We argue that grouping certain concepts together can affect the performance of the learning task. We explore the similarity between different semantic concepts and group associated concepts together to learn a few more classifiers improving the performance of video concept detection. It is further investigated if grouping of concepts in that clever way exploiting the similarity is better or if random grouping does the job. We also compare with the RAKEL framework for video concept detection. With experimentation on the TRECVID 2010 dataset and show that the clever group based classifiers outperforms random grouping of concepts and the multi-label RAKEL algorithm. We further analyze the improvements different grouping techniques bring when fused with individual concept learning for video concept detection.