Gabriel Yancy¹ (), Ajay Subramanian¹, Najib Majaj¹, Denis Pelli¹, Jonathan Winawer¹; ¹New York University

Human visual object recognition relies on a narrow, 1.5-octave-wide filter (Solomon & Pelli, 1994 & Subramanian et al., 2023). For simple stimuli like isolated letters, the center frequency of the band used grows as the ⅔ power of stroke frequency (Majaj et al., 2002, Vision Research). The ⅔ power indicates that letter recognition is scale dependent: the observer relies on higher frequencies (in cycles per letter) as object size increases. We extended the Subramanian et al. (2023, NeurIPS) noise-masking method to assess the scale dependence of natural object recognition. We presented 1,050 ImageNet images (16 categories) mixed with Gaussian noise at five strengths (SD = 0, 0.02, 0.04, 0.08, 0.16), filtered into seven log-spaced spatial frequency bands (0.2 to 14 cycles/image) to seven observers. Each stimulus was displayed for 200 ms through a square aperture of 3, 6, or 12 degrees. Participants performed a 16-way classification. Accuracy was modeled as a logistic function of noise level, with thresholds constrained to a Gaussian function of log spatial frequency. Results mirrored findings for letters: the recognition channel is about 1.5 octaves wide, and the center frequency follows a ⅔ power law of stimulus size. This is evidence that the size dependence of object recognition is not restricted to simple isolated objects but extends to a diverse range of target types.

Acknowledgements: NEI core vision grant P30EY013079 / NEI grant R01EY033628 / NEI grant R01EY027401