MIT researchers have created a comprehensive dataset of 'illusory' faces, revealing surprising differences between human and AI face detection and potentially uncovering evolutionary links to animal face recognition. The research team, led by Mark Hamilton, a PhD student in electrical engineering and computer science, has introduced a dataset of 5,000 human-labeled pareidolic images, far surpassing previous collections.
Face pareidolia has long fascinated psychologists, but it's been largely unexplored in the computer vision community. The researchers wanted to create a resource that could help understand how both humans and AI systems process these illusory faces.
The study's findings reveal significant disparities between human and machine perception of these 'fake' faces. Surprisingly, AI models initially struggled to recognise pareidolic faces in the same way humans do. It wasn't until the algorithms were trained to recognise animal faces that they showed marked improvement in detecting pareidolic faces.
Another key discovery that is highlighted: A specific range of visual complexity where both humans and machines are most likely to perceive faces in non-face objects. This phenomenon, dubbed the "Goldilocks Zone of Pareidolia," suggests that images with "just the right amount" of complexity are most likely to trigger face recognition.
The researchers developed an equation modelling how people and algorithms detect illusory faces, revealing a clear "pareidolic peak" where the likelihood of seeing faces is highest. This predicted zone was validated through tests with both human subjects and AI face detection systems.
The "Faces in Things" dataset, which forms the backbone of this study, was meticulously curated from approximately 20,000 candidate images from the LAION-5B dataset. Human annotators labelled these images, drawing bounding boxes around perceived faces and answering detailed questions about each face's perceived characteristics.