Metadata-Weighted Score Fusion for Multimedia Event Detection

We address the problem of multimedia event detection from videos captured 'in the wild,' in particular the fusion of cues from multiple aspects of the video's content: detected objects, observed motion, audio signatures, etc. We employ score fusion, also known as late fusion, and propose a method that learns local weightings of the various base classifier scores which respect the performance differences arising from the video quality. Classifiers working with visual texture features, for instance, are given reduced weight when applied to subsets of the video corpus with high compression, and the weights associated with the other classifiers are adjusted to reflect this lack of confidence. We present a method to automatically partition the video corpus into relevant subsets, and to learn local weightings which optimally fuse scores on a particular subset. Improvements in event detection performance are demonstrated on the TRECVid Multimedia Event Detection (MED) MED Test dataset, and comparisons are provided to several other score fusion methods.