In binary segmentation, should the mask be represented as a matrix of 0s and 1s where 0 is one class (background) and 1 is another class. Or should it be represented as matrix of 0s and 255s where again 0 is one class (background) and 255 is another class. Or it does not matter?
Hi Augustas!
The short answer is yes, the mask should be 0s and 1s (and, yes,
it does matter).
Specifically, when you use pytorch’s BCEWithLogitsLoss (or its
numerically-less-stable cousin BCELoss), the target (mask) you
pass in should be a floating-point tensor of probabilities that range
from 0.0 to 1.0.
When your ground-truth target has complete certainty, it would
have a value of 0.0 for the background – 0% probability of being
a foreground pixel – and a value of 1.0 for the foreground – 100%
probability of being a foreground pixel. However, by not restricting
the target to have only the values 0.0 and1.0, we admit the
possibility that the target can be probabilistic. So, for example,
a value of 0.75 would indicate that a pixel is probably a foreground
pixel – 75% likely – but could also be – 25% likely – a background
pixel.
(Using 0 and 255 – as you might have in an 8-bit black-and-white
image – won’t work. Pytorch’s BCEWithLogitsLoss requires target
values in the range [0.0, 1.0] and will give meaningless results for
values outside of this range.)
Best.
K. Frank
1 Like