Twin identification over viewpoint change: A deep convolutional neural network surpasses humans

Deep convolutional neural networks (DCNNs) have achieved human-level accuracy in face identification (Phillips et al., 2018), though it is unclear how accurately they discriminate highly-similar faces. Here, humans and a DCNN performed challenging face-identity matching task that included identical twins. Participants ( N = 87) viewed pairs of images three types: same-identity, general imposters (different identities from similar demographic groups), twin (identical siblings). The was to determine whether the showed same person or different people. Identity comparisons were tested viewpoint-disparity conditions: frontal frontal, 45-degree profile, 90-degree-profile. Accuracy for discriminating matched-identity twin-imposter imposter assessed each condition. Humans more accurate general-imposter than pairs, declined with increased viewpoint disparity between pair. A trained (Ranjan 2018) on image presented humans. Machine performance mirrored pattern human accuracy, but at above all one Human machine similarity scores compared across image-pair types. This item-level analysis ratings correlated significantly six nine types [range r 0.38 0.63], suggesting accord perception by DCNN. These findings also contribute our understanding high-resemblance faces, demonstrate performs level humans, suggest degree parity features used