UQMSG Experiments for TRECVID 2011

This paper describes the experimental framework of the University of Queensland’s Multimedia Search Group (UQMSG) at TRECVID 2011. We participated in two tasks this year, both for the first time. For the semantic indexing task, we submitted four lite runs: L A UQMSG1 1, L A UQMSG2 2, L A UQMSG3 3 and L A UQMSG4 4. They are all of training type A (actually we only used IACC.1.tv10.training data), but with different parameter settings in our keyframe-based Laplacian Joint Group Lasso (LJGL) algorithm with Local Binary Patterns (LBP) feature. For the content-based copy detection task, we submitted two runs: UQMSG.m.nofa.mfh and UQMSG.m.balanced.mfh. They used only the video modality information of keyframes and were both based on our Multiple Feature Hashing (MFH) algorithm that fuses local (LBP) and global (HSV) visual features, with different application profiles (reducing the false alarm rate v.s. balancing false alarms and misses). Due to time constraint, we were not able to improve the performance of our systems adequately on all the available training data this year for these tasks. Evaluation results suggest that more efforts need to be made to well tune system parameters. In addition, sophisticated techniques beyond applying keyframe-level semantic concept propagation and near-duplicate detection are required for achieving better performance in video tasks.