Automatic Musical Instrument Recognition in Audiovisual Recordings by Combining Image and Audio Classification Strategies

The goal of this work is to incorporate the visual modality into a musical instrument recognition system. For that, we first evaluate state-of-the-art image recognition techniques in the context of music instrument recognition, using a database of about 20000 images and 12 instrument classes. We then reproduce the results of state-of-the-art methods for audio-based musical instrument recognition, considering standard datasets including more than 9000 sound excerpts and 45 instrument classes. We finally compare the accuracy and confusions in both modalities and we showcase how they can be integrated for audio-visual instrument recognition in music videos. We obtain around 0.75 F1-measure for audio and 0.77 for images and similar confusions between instruments. This study confirms that visual (shape) and acoustic (timbre) properties of music instruments are related to each other and reveals the potential of audiovisual music description systems.