Figure 1
A screen capture of the Picasso web application after computing partial occlusion figures for various input images. The classifier is a trained VGG16 [9] network.
Figure 2
The partial occlusion map sequentially blocks out parts of the image to determine which regions are important to classification. The numbers in the header are the overall class probabilities. Brighter regions correspond to areas where the probability of the given class is high–i.e. blocking out this part of the image does not change the classification much. The tank image is in the public domain [21].
Figure 3
Saliency map for the tank. Brighter pixels indicate higher values for the derivative of “tank” with respect to that input pixel for this image. The brightest pixels appear to be in the tank region of the image, which is a good indication the model is classifying on the tank-like features.