Invariance to orientation. (a) an example of images where orientation noises are in [−20°, 20°] with uniform probability; (b) saliency map of our model for (a); (c) saliency map of Saliency Tool for (a); (d) an example of images where orientation noises are in [−30°, 30°] with uniform probability; (e) saliency map of our model for (d); (f) saliency map of Saliency Tool for (d); (g) statistics of correct detection rate vs. range of orientation noises |Δθ| in distractors (that is, when the orientation of the target is θt, the orientations of distractors vary in [θt + pi/2−Δθ,θt + pi/2 + Δθ]). For each |Δθ|, we test 20 randomly generated images. The target is marked by a red circle. The results of Saliency Tool are got by only using orientation information. Note that since orientation difference between the target and distractors is pi/2, |Δθ| should not exceed 45°.
Qi et al. BMC Neuroscience 2012 13:145 doi:10.1186/1471-2202-13-145