National Institute of Informatics, Japan at TRECVID 2012

This paper reports our experiments for four TRECVID 2012 tasks: instance search, semantic indexing, multi- media event detection, and known-item search. For the instance search task, we use the same approach as the last year's system with some modifications in quantization and fusion of query representations. The experiments show improved performance of this year's system compared to the last year's system. For the se- mantic indexing task and the multimedia event detection task, we report the experiments using NII-KAORI-SECODE framework. Our best run for the semantic indexing task using local features achieve 20.7% (MAP), ranked 6/15 groups. Especially, the run using only one local feature achieves 18.9% (MAP). As for the multimedia event detection task, we only use global features to serve as a baseline method for comparison with other complicated systems using local features and multi-modal features. I. INSTANCE SEARCH