How to use random erasing in semantic segmentation in keras imagedatagenerator?

[emma-sjwang]

Hello.

The random erasing only is performed on the image translation, right?
When we deal with segmentation tasks, there is no need to change the mask.
How can we realize it in keras?

Thanks a lot.

[yu4u]

Hi @EmmaW8

I'm not sure whether random erasing is effective for segmentation task or not, but I think it is worth trying.
You do not need to use ImageDataGenerator if it is not convenient for you. Simply use eraser function to augment training image.
It may be better to ignore losses from the erased area (this can be done by exploiting some mask that indicates masked pixels).

from random_eraser import get_random_eraser
eraser = get_random_eraser()

# load image to img
img = eraser(img)