Weighting Informativeness of Bag-of-Visual-Words by Kernel Optimization for Video Concept Detection

摘要

Bag-of-Visual-Words (BoW) feature has been demonstrated effective and widely used in video concept detection due to its discriminative ability by capturing the local information in images. In the current approaches, all the words in the visual vocabulary are treated equally for the detection of different concepts. This cannot highlight the concept-specific visual information, and thus limits the discriminative ability of BoW feature. In this paper, we propose an approach to boost the performance of video concept detection based on BoW. This is achieved by assigning different weights to the visual words according to their informativeness for the detection of different concepts. Kernel alignment score (KAS) is used to measure the discriminative ability of SVM kernels, and the visual words are weighted as a kernel optimization problem. We show that the SVMs based on weighted visual words with our approach outperform the uniformly weighting and TF-IDF weighting schemes, and the MAP for the 20 concepts from TRECVID 2009 high-level feature extraction is significantly improved.