Hamming Selection Pruned Sets (HSPS) for Efficient Multi-label Video Classification

Videos have become an integral part of our life, from watching movies online to the use of videos in classroom teaching. Existing machine learning techniques are constrained with this scaled up activity because of this huge upsurge in online activity. A lot of research is now focused on reducing the time and accuracy of video classification. Content-Based Video Information Retrieval CBVIR implementation (E.g. Columbia374) is one such approach. We propose a fast Hamming Selection Pruned Sets (HSPS) algorithm that efficiently transforms multi-label video dataset into single-label representation. Thus, multi-label relationship between the labels can be retained for later single label classifier learning stage. Hamming distance (HD) is used to detect similarity between label-sets. HSPS captures new potential label-set relationships that were previously undetected by baseline approach. Experiments show a significant 22.9% dataset building time reduction and consistent accuracy improvement over the baseline method. HSPS also works on general multi-label dataset.