KNOWLEDGE TRANSFER FROM WEAKLY LABELED AUDIO USING CONVOLUTIONAL NEURAL NETWORK FOR SOUND EVENTS AND SCENES. pdf

Authors: Anurag Kumar, Maksim Khadkevich, Christian Fügen

ESC-50 Dataset

• The dataset consists of a total of 2,000 recordings each of 5 seconds durations.
• It comes pre-divided into 5 folds.
• The training set consists of 4 out of 5 folds and the remaining 5th fold is used for testing. This is done all 5 ways and average accuracies are reported.
• The training set is used for network adaptation as well as for training linear SVMs.

ESC-50 Results

• Comparison of our proposed method with state of art methods is shown is paper
• Our method not only outperforms previous methods by a considerable margin but also outperforms human accuracy on this dataset
• Even direct representation obtained from $\mathcal{N}_S$, that is without any task adaptive training, we obtain an average accuracy of 82.8%
, compared to 81.3% human accuracy on this dataset.
• Best accuracy of 83.5% is obtained using F1 representations (with $max()$ mapping), from $\mathcal{N}_T^{I}$ and $\mathcal{N}_T^{II}$
• Class-wise results - Below we show confusion matrix for two cases. Classwise confusion matrix for all cases are available here. The file name clarifies the representation used, e.g esc50.NT_III.F1.max.png means $\mathcal{N}_T^{III}$ network, F1 representations and $max()$ function to map segment level representations to full recording level representations. The figure files here, might be visually more pleasing. All numbers have been rounded to 2 decimal places.