This paper demonstrates the framework and results from the team “Florida International University - University of Miami (FIU-UM)” in TRECVID 2015 Semantic Indexing (SIN) task [1]. Four runs were submitted, and the summary of these four runs is given as follows: • 2C M A FIU UM.15 1: MCA late fusion - Multiple Correspondence Analysis (MCA) based ranking using the MCA scores of all ten key frame (KF) features. • 2C M A FIU UM.15 2: MCA early fusion - MCA based ranking using the selected five KF features. • 2C M A FIU UM.15 3: Run 1 + Time information - MCA late fusion combined with MCA scores from frames other than key frames. • 2C M A FIU UM.15 4: MCA early fusion - MCA based ranking using the selected four KF features. In Run 1, the MCA scores from ten KF features are combined and re-ranked. For Run 3, the result from the aforementioned run (i.e., run 1) and the time information extracted from frames other than key frames are fused. In this way, we wanted to test whether the time information could help improve the results. In Run 2 and Run 4, different feature sets are combined to feed to the same baseline MCA-based model. As a result, from the submission results, Run 2 outperforms the other three runs. • Processing type: Automatic • Class: M - main, single concepts • Training type: A (only the IACC data)
