Given a still image, humans can easily think of a sound associated with this image. For instance, people might associate the picture of a car with the sound of a car engine. In this paper, we aim to retrieve sounds corresponding to a query image. To solve this challenging task, our approach exploits the correlation between the audio and visual modalities in video collections. A major difficulty is the high amount of uncorrelated audio in the videos, i.e., audio that does not correspond to the main image content, such as voice-over, background music, added sound effects, or sounds originating off-screen. We present an unsupervised, clustering-based solution that is able to automatically separate correlated sounds from uncorrelated ones. The core algorithm is based on a joint audio-visual feature space, in which we perform iterated mutual kNN clustering in order to effectively filter out uncorrelated sounds. To this end, we also introduce a new dataset of correlated audio-visual data, on which we evaluate our approach and compare it to alternative solutions. Experiments show that our approach can successfully deal with a high amount of uncorrelated audio.
The documents contained in these directories are included by the contributing authors as a means to ensure timely dissemination of scholarly and technical work on a non-commercial basis. Copyright and all rights therein are maintained by the authors or by other copyright holders, notwithstanding that they have offered their works here electronically. It is understood that all persons copying this information will adhere to the terms and constraints invoked by each author’s copyright. These works may not be reposted without the explicit permission of the copyright holder.