Skip to main content
. 2020 Oct 7;7(10):200595. doi: 10.1098/rsos.200595

Table 2.

Performance of the CNNs and humans on the four matching tasks: match/mismatch per cent correct. AUC is area under the curve across all tests combined, calculated using perfcurve in Matlab. Human AUC data comes from the average response to each face pair and therefore reflects ‘the wisdom of the masses'.

1 2 3 4 5 6 human
Kent 99.5/95 83/95 99.5/80 98.5/95 77.5/100 83/100 77.6/63.8
Models 100/97.8 91.1/95.6 100/77.8 100/97.8 75.6/97.8 93.3/97.8 67.8/77.2
Make-up 100/98.6 80.2/100 100/87.7 98.6/99.3 79.6/100 98.6/87.1 73.9/79.8
Dutch 97.9/100 41.7/97.9 100/83.3 97.9/100 35.4/97.9 29.2/89.6 70.5/74.9
AUC 0.9997 0.9824 0.9974 0.9979 0.9823 0.9859 0.9671