Purchase this article with an account.
Alan Chauvin, Daniel Fiset, Catherine Ethier, Karine Tadros, Martin Arguin, Frederic Gosselin; Spatial frequency streams in natural scene categorization. Journal of Vision 2005;5(8):603. doi: https://doi.org/10.1167/5.8.603.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
We used the Bubbles method (Gosselin & Schyns, 2001) to examine the effective use of spatial frequencies through time in natural scene categorization. Two observers (C.E and K.T) categorized a total of 8640 dynamic stimuli (6 deg2 of visual angle*180ms) composed of one of 720 natural scenes from six categories (beach, city, mountain, forest, highway and landscape). Each of our stimuli was composed of 18 frames, made from the dot product of the Fourier spectrum of a scene with 2D white Gaussian noise convolved with a Gaussian function (Std's = 0.08 of the Nyquist frequency and 79 ms). We performed a linear regression on reaction times and sampling noise. The resulting classification image shows the use of different spatial frequencies across the 18 frames composing every animation. We conducted a one tailed Z-score analysis based on random field theory (Chauvin et al, submitted) in order to reveal the use of spatial frequency as a function of time. The classification image (Z > 3.8, p < 0.01) reveals the use of three narrow bands of spatial frequencies across time. Low frequencies (1 c/dg) are first to reach signifiance (between 10 and 90 ms), followed by mid frequencies (6 c/dg, significant between 30 and 120 ms) and finally higher frequencies (12 c/dg, significant between 70 to 100 ms). Superficially, this result corroborates the coarse-to-fine hypothesis (Parker, Lishman, & Hughes, 1992) of natural scenes categorization. It allows, however, a much finer analysis of the information subtending the first moments of visual categorization.
This PDF is available to Subscribers Only