Saivydas Villani¹ (), Guido Maiello¹; ¹University of Southampton

We’ve all had the experience of pointing out a “face” in a plug socket, cloud, or car bumper, only to be met with a puzzled look and a polite “I don’t see it.” This everyday disagreement raises a deeper question: does the strength of pareidolic impressions vary reliably across images and observers? And what does this reveal about the mechanisms that drive illusory face perception? To answer this, we curated a subset of the FacesInThings dataset, a pareidolia image set with Easy/Medium/Hard human annotations (from a single rater) for the difficulty of perceiving the illusory face (Hamilton et al., ECCV, 377–395). Our final test set contained 25 stimuli per difficulty level, plus 25 independent control images. We presented these stimuli to seven human observers and two computational face detection models (RetinaFace with MobileNet and ResNet‑50 backbones) fine-tuned on the FacesInThings training set to identify pareidolic faces. Humans and models were tasked with detecting and localizing pareidolic faces and rated the “faciness” of the pareidolic images. FacesInThings difficulty levels systematically affected human behaviour. The probability of detecting a pareidolic face as well as the “faciness” ratings decreased from Easy to Hard (p<.001), while reaction times increased (p<.001). Face detection models also followed this pattern, correctly identifying the pareidolic faces with decreasing confidence as difficulty level increased (p<.001 for both models). Additionally, the strength of pareidolia (the “faciness” ratings) varied across individual images (average between-participant agreement r̄=.70, p<.001) and participants (split-half reliability r=.95, p<.001). These data demonstrate that the strength of pareidolic impressions varies markedly and reliably across stimuli and individuals. Further, computational face detection models can also be trained to exhibit pareidolia-like responses, serving as a scalable, objective measure of pareidolic sensitivity, while also allowing us to model individual‑ and image‑level differences in illusory face perception.

Acknowledgements: This research was funded by a South Coast DTP studentship awarded to Saivydas Villani