TUB @ MediaEval 2012 Tagging Task: Feature Selection Methods for Bag-of-(visual)-Words Approaches.
Sebastian Schmiedeke, Pascal Kelm, Thomas Sikora · MediaEval · 2012
This paper describes our participation in the Genre Tagging Task of MediaEval 2012, which aims to predict the videos’ category label. In last year’s participation, we performed experiments with bag-of-words (BoW) approaches in which dierent constellations in respect of modalities, features, and methods were investigated. This year, we focus on feature selection methods to improve the classication performance in terms of mean average precision (mAP) and classication accuracy (CA). We investigated the eectiveness of selection methods based on scores calculated using mutual information (MI) or term frequency (TF) and the eectiveness of transformation methods like the principle component analysis (PCA).