TUT wins two of the top-five places in automated wildlife-monitoring challengeTUT's Audio Research Group (TUT-ARG) recently participated in an international challenge on Bird Audio Detection. Among the 30 teams competing in the challenge, TUT-ARG's submissions came in 2nd and 5th places.
According to BirdLife International, 13% of the bird species in the world are threatened by extinction. Almost half of the threatened species reside on islands, particularly on oceanic islands far from land. Remote wildlife-monitoring for birds is therefore crucially important for recording the bird population densities and assessing the risk of extinction.
Many wildlife-monitoring projects around the world rely on heavy manual labor. Unattended remote recorders collect huge amounts of audio or video data, which must then be evaluated by human listeners for the presence of birds. Using an automated process for this task would save valuable time and data storage.
In order to stimulate research on bird audio detection, Queen Mary University of London, University of Toulon, University of Crete, and University of Salford recently joined forces to organize a bird audio detection research challenge. The aim of the challenge was to design automated systems with the highest accuracy for detecting birds from their sounds. In the challenge, the competitors were provided with three different datasets. Each dataset contained 10-second recordings from diverse physical environments where the birds are likely to be located. The recordings for two of the datasets were labeled by human annotators to reflect the presence of a bird in the recording. These datasets were used by the competitors to implement their machine learning solutions, which were evaluated on the third dataset.
Among the 30 teams competing in the challenge, the submission from Emre Cakir (photo), a PhD student in TUT's Audio Research Group (TUT-ARG), placed second and won the Judges' award. The proposed method utilized a combination of deep learning techniques, which is currently the hottest trend in machine learning. The proposed method was able to detect the bird sounds with 88.5% accuracy, which came close to the winning 88.7% achieved by the Austrian Research Institute for Artificial Intelligence (OFAI). Another method proposed by Sharath Adavanne, who is also a PhD student in TUT-ARG, came in fifth with 88.1% accuracy. Detailed information on the submitted methods and the results can be found on the Bird Audio Detection Challenge website.