Fig 3. Systematic occlusion of parts of the image.

A) Examples where we occluded different portions of the scene, and visualized how the classifier output for the correct class changed (before the softmax activation function). Images were occluded by a gray patch of 128x128 pixels, sliding across the image in 32 pixel steps. Importance is defined as the relative change in activation after occluding that part of the image (compared to the activation of the ‘original’ unoccluded image) and is computed as follows: original activation—activation after occlusion / original activation. This example is for illustrative purposes only; maps vary across exemplars. B) The relative change in activation (compared to the original image), after occluding pixels of either the object or the background, for the different conditions (segmented, congruent, incongruent). For each image, importance values of the objects and backgrounds were normalized by dividing them by the activation for the original image, resulting in the importance ratio. Error bars represent the bootstrap 95% confidence interval. Non-significant differences are indicated with a solid line below the graph.