Table 7 Average ranking test performance (over 30 classes), when late fusing individual classifiers trained on sub-feature families of Audio and Visual features (14 many subfamilies), where late fusion is achieved by learning on the validation data (no learning for SUM). We observe a significant boost in rec@99, in particular via random forests

From: On using nearly-independent feature families for high precision and confidence

  Prec.
99 % 95 % 90 % Max F1
Random forests (200 trees) 0.53 0.65 0.73 0.84
Perceptron committee (40) 0.50 0.64 0.72 0.83
Linear SVMs, C=10 0.49 0.63 0.72 0.83
SUM 0.47 0.59 0.67 80