Correspondence between audio and visual deep models for musical instrument detection in video recordings

Abstract

This work aims at investigating cross-modal connections between audio and video sources in the task of musical instrument recognition. We also address in this work the understanding of the representations learned by convolutional neural networks (CNNs) and we study feature correspondence between audio and visual components of a multimodal CNN architecture. For each instrument category, we select the most activated neurons and investigate existing cross-correlations between neurons from the audio and video CNN which activate the same instrument category. We analyse two training schemes for multimodal applications and perform a comparative analysis and visualisation of model predictions.