This paper presents the framework and results of the team "Florida International University - University of Miami (FIU-UM)" in TRECVID 2012 Semantic Indexing (SIN) task [3] [13]. Four runs of the SIN results were submitted, and the summary of the four runs is as follows: F A FIU-UM-1-brn 1: Fusion of the results generated from three models, corresponding to the rest of the three runs. F A FIU-UM-2 2: SMR+KF+CAN - Subspace Modeling and Ranking (SMR) using the Key Frame-based low-level features (KF). The Concept Association Network (CAN) is applied to the ranking results to improve some poor detected concepts according to their relationship to other concepts. • F A FIU-UM-3-brn 3: MCA+KF+SF+CAN - Multiple Correspondence Analysis (MCA) based ranking using KF features and shot-based features(SF). CAN is applied to the ranking results of this round as well. F A FIU-UM-4 4: LR+KF+CAN - Logistic Regression (LR) using KF features. CAN is applied to the ranking results of this round as well. In Runs 2, 3, and 4, each of them uses a different learning algorithm to train the model and predict testing instances. KF features are used in all these runs, but SF features are used only in Run 3. Additional training labels provided by NIST are also used in Run 3 (called "brn" in the name) as a trial. The Concept Association Network (CAN) is applied to all these runs to utilize the correlation between the concepts to improve the concepts with poor performance by the concepts with good performance. Finally, the results of these three runs are fused together to generate Run 1 as the best run. From the submission results, Run 1 does perform the best among all the four runs.
