occlusion dark spots

[rajasekharponakala]

Hi,

could you help me to get the blackout dark spots on the generated images, in the way you used in the paper (Figure 8).

Here is what my images look alike:

can i get the shorter dark rectangles, could you reproduce the size that we have to use for occl_size with the arg.parse(). Thanks.

[peikexin9]

Sorry @rajasekharponakala, I do not quite remember the exact size. From just the look, I suspect it should be something around 5x5?

[rajasekharponakala]

@peikexin9: Thanks. For example, is the syntax correct? python gen_diff.py -occl_size 5x5 blackout 1 0.1 10 100 50 0

[peikexin9]

@rajasekharponakala try something like -occl_size (5,5)?

[rajasekharponakala]

@peikexin9: As the paper discuss on domain-specific constraints, can we consider them as high-dimensional spaces, which are linear in nature and causing adversarial examples?

[peikexin9]

@rajasekharponakala I would think so. There must be a space linear to these transformations that the NN should be invariant of (while it is actually not thus resulting in the adversarial example).