image_and_masks contains black png images

[hxngillani]

why there are black .png images in mask folders ?

[ianhi]

Hi @hxngillani those are where the mask data are stored. They show up as black when you look at them because they only hold the class data for each pixel. So the only values in the file are 0,1,2. Since these are all very low values the PNG will appear black when you try view it with an image viewer. We do this because it's easier to store the data this way than to store the colors of each class. So we store them as 0,1,2 and then convert the classes to colors when we want to display them. The code that does this when you load a premade mask is here:

AC295-final-project-JWI/lib/labelling.py

Line 213 in 4f1831a

self.colors = np.vstack([[0,0,0],plt.get_cmap(self.colors)(np.arange(len(classes)))[:,:3]])

AC295-final-project-JWI/lib/labelling.py

Lines 378 to 380 in 4f1831a

	idx = self.class_mask != 0
	c_overlay = self.colors[self.class_mask[idx]]*255*self.overlay_alpha
	array[idx] = (c_overlay + self.img[idx]*(1-self.overlay_alpha))

[hxngillani]

thanks you ianhi :)

[MBelkhechine]

Hi,

thank you for explaining this question.

Would you please tell me, how you are storing the class indexes? Is it in the B value of the RGB color scheme?

[ianhi]

@MBelkhechine good question. They are actually stored as black and white images so the arrays have shape (450, 450). You can check this by opening one up using skimage.io.imread:

from skimage import io
arr = io.imread('data/image_and_masks/train_masks/train/0_159350906_z_-2.5.png')
print(arr.shape)

[MBelkhechine]

Thank you for your prompt response.

I had a closer look at the masks provided in the dataset and it seems that the class indexes are stored in all the RGB color space dimensions e.g. Pixels labelled as cell2 will have this tuple value rgb(2,2,2).