Clarify GT format

[bertsky]

Please help me understand this:

Ground truth format

Lables for each pixel are identified by a number. So if you have a
binary case, n_classes should be set to 2 and labels should
be 0 and 1 for each class and pixel.

In the case of multiclass, just set n_classes to the number of classes
you have and the try to produce the labels by pixels set from 0 , 1 ,2 .., n_classes-1.
The labels format should be png.

If you have an image label for a binary case it should look like this:

Label: [ [[1 0 0 1], [1 0 0 1] ,[1 0 0 1]],
[[1 0 0 1], [1 0 0 1] ,[1 0 0 1]] ,
[[1 0 0 1], [1 0 0 1] ,[1 0 0 1]] ]

This means that you have an image by 3*4*3 and pixel[0,0] belongs
to class 1 and pixel[0,1] belongs to class 0.

First, that's just not true: the array shown is not 3*4*3, but 3x3x4. Next, pixel[0,0] then is not 1, but [1 0 0 1].

Now, when you say The labels format should be png here, I take it GT is supposed to be presented as pseudo/indexed-color PNG files of the same resolution as the input image and with color values for classes. Is that correct so far?

Then why does the example array have 3 (not 2) dimensions?

Maybe it would also help to show an actual image / label-image pair example ...

[ss756]

Would be helpful if some could clarify the structure for constructing the mask for the images