Accurate recognition and discrimination of complex visual stimuli is critical to human decision making in medicine, forensic science, aviation, security, and defense. This study highlights the sufficiency of redundant low-spatial and low-dimensional information for visual recognition and visual discrimination of 3 large-scale natural image sets.
Humans can see through the complexity of scenes, faces, and objects by quickly extracting their redundant low-spatial and low-dimensional global properties, or their style. It remains unclear, however, whether semantic coding is necessary, or whether visual stylistic information is sufficient, for people to recognize and discriminate complex images and categories.
In 2 experiments, we systematically reduce the resolution of hundreds of unique paintings, birds, and faces, and test people’s ability to discriminate and recognize them.
We show that the stylistic information retained at extremely low image resolutions is sufficient for visual recognition of images and visual discrimination of categories. Averaging over the 3 domains, people were able to reliably recognize images reduced down to a single pixel, with large differences from chance discriminability across 8 different image resolutions. People were also able to discriminate categories substantially above chance with an image resolution as low as 2×2 pixels.
We situate our findings in the context of contemporary computational accounts of visual recognition and contend that explicit encoding of the local features in the image, or knowledge of the semantic category, is not necessary for recognizing and distinguishing complex visual stimuli.
[Keywords: visual recognition, visual discrimination, ensemble, gist, perceptual expertise]
