Cascade of Boolean detector combinations

EURASIP Journal on Image and Video Processing

Table 1 Results in laughter detection task

	acc.	\(F_{1}^{\text {sp}}\)	\(F_{1}^{\text {lg}}\)	v.f. %
(l_A≥θ_A)	95.1	.944	.927	0%
(l_V≥θ_V)	84.4	.818	.795	100%
BOA cascade of (15), N=1	96.0	.966	.958	11%
BOA cascade of (15), N by BOATS	96.9	.972	.955	33%
C5.0 tree	96.7	.965	.948	23%
Boosted C5.0 forest	97.3	.978	.958	≈ 100%
[53]	92.7	.943	.905	100%
[54]^a	91.7	.932	.893	100%
[55]^b	96.9	.973	.963	100%

Comparison of laughter detectors on MAHNOB laughter data. The used measures of performance are the overall accuracy, F₁ -scores for both speech (\(F_{1}^{\text {sp}}\)) and laughter (\(F_{1}^{\text {lg}}\)), and percentage of video clips, the classification of which utilized also visual features (v.f.). The BOA detectors are used at the operating point α of the highest accuracy on training set
^aComparison with [54] is not directly comparable, as the classifier in [54] is trained with another dataset
^bResults of [55] are with 15 speakers while the other authors use 22 speakers in their tests