Evaluation datasets published for Tasks 2-4

Toni Heittola

Evaluation datasets for Tasks 2, 3 and 4 are now published, and the download links can be found in the download page and task description pages. The baseline system is also updated to support these evaluation datasets.

Evaluation dataset for Task 1 will be published in one week.

Make sure you pay attention to the recent updates on the development datasets:

Task 2

A minor bug was found in the the development dataset after initial release. A few background audio signals were found to contain also target sound events (babycry), affecting the evaluation result minimally. Make sure you update your dataset by using the provided script. Detailed instructions how to use the script can be found here.

See more in DCASE Forum post

Task 3

Reference annotations in the Task 3 development dataset were verified after initial release as it was noticed that they did not always correspond to audible sound events. To address this issue, a verification process was implemented. Make sure you update your development dataset to version 2. This verified dataset will be used officially for the DCASE challenge.

See more in DCASE Forum post

Task 4

A few items in the strong label annotations were found to have class label ("car horn") not used in the challenge.

Please update file groundtruth_strong_labels_testing_set.csv.

To make sure the development dataset is downloaded correctly from YouTube, two methods are provided:

  • scripts to collect file durations and compare them to correct durations for training set and testing set.
  • Sharing a direct-download-link containing the clips of the training and testing sets "own" by the organisers: Training set (Password: DCASE_2017_training_set) and Testing set (Password: DCASE_2017_testing_set).

See more in DCASE Forum post