A Music Video Information Retrieval Approach to Artist Identification

We propose a cross-modal approach based on separate audio and image data-sets to identify the artist of a given music video. The identification process is based on an ensemble of two separate classifiers. Audio content classification is based on audio features derived from the Million Song Dataset (MSD). Face recognition is based on Local Binary Patterns (LBP) using a training-set of artist portrait images. The two modalities are combined using bootstrap aggregation (Bagging). Different versions of classifiers for each modality are generated from subsamples of their according training-data-sets. Predictions upon the final artist labels are based on weighted majority voting. We show that the visual information provided by music videos improves the precision of music artist identification tasks.