Twin Identification over Viewpoint Change: A Deep Convolutional Neural Network Surpasses Humans

Area of Science:

• Computer Vision
• Cognitive Science
• Biometrics

Background:

• Deep convolutional neural networks (DCNNs) demonstrate human-level accuracy in face identification tasks.
• The ability of DCNNs to discriminate between highly similar faces, such as identical twins, remains less understood.
• Human face perception capabilities are well-established but can be challenged by subtle variations and high-resemblance individuals.

Purpose of the Study:

• To compare the performance of humans and a DCNN in a challenging face-identity matching task involving identical twins.
• To assess how viewpoint disparity affects the accuracy of both humans and the DCNN in discriminating face identities.
• To investigate the correlation between human and DCNN similarity judgments for various face pair types.

Main Methods:

• A face-identity matching task was administered to 87 human participants and a DCNN.
• Image pairs included same-identity, general imposter, and identical twin imposter categories.
• Comparisons were conducted under three viewpoint-disparity conditions: frontal-to-frontal, frontal-to-45° profile, and frontal-to-90° profile.

Main Results:

• Human accuracy was higher for general imposters than twin imposters, declining with increased viewpoint disparity.
• The DCNN mirrored human accuracy patterns, performing at or above human levels in most conditions.
• Human and DCNN similarity scores showed significant correlations across six of nine image-pair types.

Conclusions:

• DCNNs exhibit strong performance in discriminating high-resemblance faces, comparable or superior to human capabilities.
• The findings suggest that DCNNs utilize features for face discrimination that align with human perceptual strategies.
• This research advances understanding of DCNNs in challenging biometric applications and highlights their potential in forensic and security contexts.