The view tolerance of human identity recognition depends on horizontal face information.

This study investigates which visual information enables humans to recognize facial identity across different viewpoints, a key unresolved question in vision science. Participants completed an identity recognition task using faces rotated across a range of yaw angles and filtered to retain specific orientation ranges of visual information. Regardless of viewpoint, human performance consistently relied on horizontal facial information. To understand why, we used model observers to assess the identity information physically available in the images. A view-selective model, which matched identities within the same viewpoint, indicated that diagnostic identity cues shift from predominantly horizontal in frontal views to more vertical in profile views. In contrast, a view-tolerant model, which matched identities across different viewpoints, revealed that horizontal information provides the most stable and reliable identity cues across views. Furthermore, horizontal facial information best predicted the average appearance of a face across viewpoints, supporting its role in forming stable identity representations. These findings suggest that view-tolerant face representations are acquired through exposure to the stable statistical properties of faces primarily conveyed by horizontal information. By specifying the spatial information underlying recognition across viewpoints, the study offers valuable empirical constraints for the development of theoretical and computational models of face recognition.