Columbia University TRECVID 2007 High-Level Feature Extraction.
Shih‐Fu Chang, Wei Jiang, Akira Yanagawa, Eric Zavesky · 2007
High-level feature extraction • A COL base6 T2 6 (R6): train on TRECVID2007 development data, multi-parameter models with average fusion across several modalities (referred to as target baseline) • A COL xd5 T5 5 (R5): train on TRECVID2007 development data and TRECVID2005 development data using feature replication, multi-parameter models with average fusion across several modalities (referred to as xd or replication) • A COL bcrf base4 T7 4 (R4): trained with contextual scores from TRECVID2005 based models (using columbia374) and with TRECVID2007 development (referred to as BCRF) then average-fused with target baseline • A COL bcrf xd base3 T14 3 (R3): average fusion of BCRF, target baseline, and xd replication • A COL bcrf xd base col3742 T16 2 (R2): average fusion of BCRF, target baseline, xd replication, and columbia374 • A COL best of all1 T17 1 (R1): choose best-performing classifier for each concept over a validation data subset from all above submissions • col374: not submitted, but publicly available model (see [16]) set covering 374 concepts in the LSCOM