lnugraha / trimap_generator Goto Github PK

There might be a tiny "clump" of pixels which share the same pixel intensity as foreground.
The trimap generator will dilate those non-dominant regions.
An error-handler (either reject or convert those clumps as background) is required to anticipate this issue (see the attached image).

My temporary proposed solution is to utilize clustering method to detect the most dominant foreground.

Hello,

Thank you for your work, it's very useful.

I have develop a solution for the optimization part with cython.
I will share you the file to paste in folder where the trimap module is located. You need to dezip it.
And in trimap_module.py add theses lines :
from trimap.faster_remake import set_value

And replace (in trimap function) this :
for i in range(0,row):
for j in range (0,col):
if (remake[i,j] != 0 and remake[i,j] != 255):
remake[i,j] = 127
by this :
remake = np.array(set_value(remake, 127))

You can change the value if you want.

Hope that's help.

Have a good day and enjoy :)

faster_remake.cpython-36m-x86_64-linux-gnu.zip

Make an example, like slic algorithm in scikit-image.

Hi Inugraha,

It seems that you dilate the input binary mask to generate the trimap. But what if the input mask is not confident enough to be sure foreground? E.g., masks generated by Mask RCNN segmentation sometimes extend outside the ground truth objects. So perhaps shrink the input mask a bit (e.g., erosion operation) as sure foreground before dilation?

TODO: design an abstract class that can decide whether erosion or dilation will be used for the foreground. The derived classes deploy erosion or dilation with a kernel size decided priorhand

hi, lnugraha
I do not find strongly-connected-component algorithms in your code. Can you show me paper about how to use strongly-connected-component algorithms in trimap?
thanks