The NTU Toolkit and Framework for High-Level Feature Detection at TRECVID 2007

In TRECVID 2007 high-level feature (HLF) detection, we extend the well-known LIBSVM and develop a toolkit specifically for HLF detection. The package shortens the learning time and provides a framework for researchers to easily conduct experiments. We efficiently and effectively aggregate detectors of training past data to achieve better performances. We propose post-processing techniques, concept reranking and temporal filtering, to exploit inter-concept contextual relationship and inter-shot temporal dependency. The overall improvement is 46% over that by our baseline in terms of infMAP . We briefly summarize our six submitted runs in this abstract. The run (runid: A nt20Giants 6) adopts multiple low-levels features (all visual features), SVM models, ensemble bagging classifier, and multimodal fusion. We take this setting as our baseline. We then experiment with post-processing methods and the leverage of classifiers using past data. The proposed post-processing framework is firstly applied to the baseline to obtain a new run (runid: A ntMonster 4). in terms of infMAP , this new run improves 16.7% over the baseline The runs, A ntTank05 1 and A ntTransformer 5, aggregate classifiers of using past data by averaging and weighted averaging their results, respectively. The results of these two runs, A ntTank05 1 and A ntTransformer 5, are respectively 17.3% and 25.0% higher than that of A ntMonster 4. Based the observation of our experimental results, we conclude that post-processing and using past data are helpful to improve HLE detection. Table 1. Description of each submitted run HLF Run infMAP Description A nt20Giants 6 0.0599 BASELINE: 20 bagging classifiers, multi-modal average fusion. A ntMonster 4 0.0699 bagging classifiers, weighted fusion, post-processing. A ntTank05 1 0.0820 bagging classifiers, weighted fusion, average aggregation, post-processing. A ntTransformer 5 0.0874 bagging classifiers, weighted fusion, weighted aggregation, post-processing. A ntReranking 3 0.0756 bagging classifiers, weighted fusion, average aggregation, reranking. A ntFiltering 2 0.0787 bagging classifiers, weighted fusion, average aggregation, filtering after reranking. ? This work is primarily supported by the National Science Council of Taiwan, R.O.C., under contracts NSC95-2622E-002-018 and NSC96-2622-E-002-002.