Figure 1.

Performance of a linear classifier feed with amplitude-only information to classify face and vehicle images used in Crouzet et al. (2010). After having resized images to 256 × 256, they were passed trough a Hamming window function to remove boundary artifacts. The amplitude of the Fourier transform was then computed on each image. The resulting distribution of frequencies were divided into four bins of orientation (horizontal, vertical, and the two obliques), each one covering 45°. The distribution of frequencies for each orientation was then encoded by 20 points, resulting in 80 values representing the global features of each image to feed the classifier. A linear SVM was then trained on half of the images (50% faces, 50% vehicles) and tested on the other half. After 1000 cross-validations with random train and test subsets, we computed the mean performance of the classifier to correctly classify an image in its class. The error bars correspond to the SD. The computing of the global and unlocalized features was based on the MATLAB script AverageAndPowerSpectrum.m (Torralba and Oliva, 2003), and the classification was done using the LIBSVM 2.9-1 for MATLAB (http://www.csie.ntu.edu.tw/~cjlin/libsvm/).