Some questions about the implementation

[palatos]

Thank you for this great and straight forward implementation of this technique!
I was going through the code to understand a bit more of how it works, and I'm curious about some parts:

There seem to be two blurred images:

blurred_img1 = cv2.GaussianBlur(img, (11, 11), 5)
blurred_img2 = np.float32(cv2.medianBlur(original_img, 11))/255
blurred_img_numpy = (blurred_img1 + blurred_img2) / 2

Futher ahead only blurred_img2 is used for training the mask:

img = preprocess_image(img)
blurred_img = preprocess_image(blurred_img2)
mask = numpy_to_torch(mask_init)

But for heatmap generation we use the average of the two blurs:
save(upsampled_mask, original_img, blurred_img_numpy)

I was a bit confused by this. Is there a reason why it is done in this way?