How does same class different instances distinguished?

I logged out sizes of output:

INFO 12.22 11:31:34 demo.py:65: boxes: (15, 4)
INFO 12.22 11:31:34 demo.py:66: masks: torch.Size([15, 720, 1280])
INFO 12.22 11:31:34 demo.py:67: scores: (15,)

Seems every mask contains more than one instance.

This got a situation that, some mask didn’t have mask related with it:

how to remove those separate masks that does instance or does have box related with it?

[dbolya]

This is the kind of output you get when you don’t crop the resulting masks. The model is trained with cropping, so it leverages that fact and doesn’t bother segmenting out instances that are on the other side of the image for instance.

What it’s learned is that it only needs to supress instance boundaries when boxes overlap, which is exactly what you’re seeing in that image (notice the masks of different colors stop at the boundaries between objects). Essentially, the model doing as little instance segmentation as it can get away with and it lets cropping do the rest. Naturally, there will be errors but that’s life ¯\_(ツ)_/¯

Make sure you crop before you visualize.

[seawater668]

@dbolya

How to label a single category with labelme? After looking at the user-defined data set, there is no annotation naming method. For example, there is only one category in a graph: cat, cat, cat or CAT1, cat2, Cat3 or cat-0, cat-1, cat