Abstract
Whether a scene is to be classified at the superordinate level as man-made or natural, the basic level as a particular outdoor scene (e.g., mountain), or as a specific example of an outdoor scene at the subordinate level (e.g., the Himalayas) will modify the visual information required. To approach these issues, we focussed on the superordinate categories, Oman-made and Onatural, and implemented an ideal observer simulation using 3200 images of scenes (from Oliva & Torralba, 2001) to derive a benchmark of the information available for these categorizations (A). Stimuli were the scenes with phase noise introduced at randomly selected spatial frequencies and orientations in Fourier Space. The density of the noise was adjusted to maintain classification performance at 75% correct. The same method was applied to human observers, for the same 75% accuracy level to derive D, the information Diagnostic of these categorizations. We then computed the relative efficiencies (E = D/A) of the scene classifications. Stimuli reconstructed from the diagnostic Fourier coefficients reveale d the information (both in terms of bandwidths and orientations) that characterized the human biases.