Weighting informativeness of bag-of-visual-words by Kernel optimization for video concept detection

Bag-of-Visual-Words (BoW) feature has been demonstrated               e®ective 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 dif-               ferent concepts. This cannot highlight the concept-speci¯c               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 di®erent weights to the               visual words according to their informativeness for the de-               tection of di®erent 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 weight-               ing and TF-IDF weighting schemes, and the MAP for the 20               concepts from TRECVID 2009 high-level feature extraction               is signi¯cantly improved.

Document Doi Bibtex

Permalink: http://www.eurecom.fr/publication/3245