Mask_rcnn: minimize_mask in utils.py returns wrong values

[AravindhKuppusamy]

The mini mask is converted to all zeros(False) on the scipy imresize function when the mask contains only 1's.(ie., if the bounding box rectangle doesn't contain even one background pixel).

m = scipy.misc.imresize(m.astype(float), mini_shape, interp='bilinear')
mini_mask[:, :, i] = np.where(m >= 255, 1, 0)

[keven4ever]

@AravindhKuppusamy good finding! But will it be possible that there is no bg pixel in bounding box? In this case, the mask is a rectangle which is the bounding box, right?

[ChunHsinWang]

I've seen someone proposed a fix in one of the issues, but cannot find a reference to it. Simply replace the line with the following:
m = scipy.misc.imresize(m.astype(float)*255, mini_shape, interp='bilinear', mode='F')
mini_mask[:, :, i] = np.where(m >= 128, 1, 0)

You will probably need to change the them in expand_mask as well

[waleedka]

Just pushed a fix for this issue. I'm closing it, but please let me know if you still encounter an edge case that I might've missed.

[John1231983]

@waleedka: i think it has some bug in somewhere because i used it and got stop program during training. I reported in the other thread

[AravindhKuppusamy]

I am still facing the same problem even with skimage transform resize function. Still getting 0s if the mask covers entire bounding box!

[waleedka]

Okay, reopening this then. I can't reproduce the problem on my end, so can someone send me a stack trace that shows the error?

[AravindhKuppusamy]

Now, none of the mini-masks contain any class pixels other than background after the push. You can just try running inspect_data.ipynb file. In the part "add augmentation and mask resizing", you wont see any mask associated with class.

[nhanteh]

I have exactly the same issue, I will post here if I find the reason

[waleedka]

Please try the latest update as of 15 minutes ago. There was a separate thread that identified a bug in skimage.transform.resize. I don't know if it's related, but it happened around the same time so maybe it is.

Also, I just ran the latest code on colab, and I do see masks in the train_shapes notebook. Please let me know if this is still an issue for you.

[dmitrykudinov]

@waleedka I think it has to do with the need for preserve_range=True in the skimage.transform.resize call. From https://github.com/scikit-image/scikit-image/issues/2702

PS: I am on

>>> skimage.__version__
'0.13.1'

[waleedka]

I think we addressed two issues here. The original at the top, about rectangular masks being zeros, which was fixed (see my first comment above). Then there was a side effect mentioned later and it was fixed as well in a later commit.

So my understanding is that this issue is fixed. If there is still a problem, it's likely to be a different one than the ones addressed here, and I would prefer to track it in a separate issue. So I'm closing this and will monitor other issues in case there are still edge cases we missed.