Yong Xu, Qiuqiang Kong, Wenwu Wang and Mark Plumbley won the 1st prize in Task 4, ‘large-scale weakly supervised sound event detection for smart cars’, Subtask A, ‘audio tagging’ in the IEEE AASP Challenge on Detection and Classification of Acoustic Scenes and Events (DCASE 2017). The DCASE challenge constitutes the most important challenge in the non-speech audio domain. It is organised by Tampere University of Technology, Carnegie Mellon University and INRIA and sponsored by Google and Audio Analytic. Because of its unique standing, the best players in the field participate such as CMU, New York University, Bosch, USC, TUT, Singapore A*Star, Korean Advanced Institute of Science and Technology, Seoul National University, National Taiwan University and CVSSP.

‘Making Sense of Sounds’ postdoctoral fellow Yong Xu and and PhD student Qiuqiang Kong contributed equally to the winning system.

Their best system came first according to the deciding F1 metric with 55.6%, outperforming the next best competitor by a notable 3%. On the other metrics their best system scored 6th (precision) and 3rd place (recall).

These results also secured them the winning position in the teams ranking, in which only the best system of each contributor is counted.

For Subtask B ‘Sound event detection’ the best system of the same team grabbed third place according to the deciding ER metric and fourth according to the complementary F1 metric.

They came second in the teams ranking here.

A detailed overview over all results can be found here and details of the submitted systems in the accompanying technical report:

Surrey-Cvssp System for DCASE2017 Challenge Task4 (PDF)

 

Christian Kroos and Mark Plumbley achieved 14. place (out of 34) with their best system in Task 3 ‘sound event detection in real life audio’ according to the deciding ER metric and came first according to the complementary F1 metric! Their contribution was one of the very few systems that did not employ a variant of a deep neural net (DNN) and used neuroevolution for the classification and detection task. Despite of the parsimonious networks that are typically the output of the artificial evolution process, it appears to be a serious competitor to the much larger DNNs.

The results of this task can be found here and details of the submitted systems in the accompanying technical report:

Neuroevolution for Sound Event Detection in Real Life Audio: a Pilot Study (PDF)

  ⤧  Previous post Sobieraj et al. (2017a)