Frequency-dependent auto-pooling function for weakly supervised sound event detection

EURASIP Journal on Audio, Speech, and Music Processing

Table 6 The results of MIL and source-separation-based methods

Method		Parameters	Audio tagging			Sound event detection			Error rate
		10 k	F-score	AUC	mAP	F-score	AUC	mAP	ER	D	I
MIL	Attention [20]	54.15	0.671	0.923	0.723	0.341	0.861	0.348	1.574	0.885	0.689
	TALNet [19]	94.06	0.646	0.911	0.687	0.397	0.849	0.390	1.339	0.865	0.474
Source	VGG-GWRP [24]	58.76	0.572	0.923	0.635	0.429	0.803	0.372	1.991	0.780	1.210
separation	VGG-AP	58.76	0.538	0.909	0.639	0.352	0.823	0.362	1.886	0.844	1.061
-based	VGG-FAP	58.76	0.590	0.923	0.672	0.407	0.848	0.385	1.776	0.823	0.952
	DDC-GWRP	28.84	0.626	0.931	0.689	0.468	0.808	0.404	1.850	0.813	1.037
	DDC-AP	28.84	0.573	0.919	0.684	0.382	0.845	0.398	1.831	0.853	0.978
	DDC-FAP	29.10	0.633	0.931	0.719	0.446	0.868	0.427	1.689	0.845	0.844