get the dist, points, scores from the def non_maximum_suppression_inds function in the config ?

[Nal44]

Hi Stardist team,

I am wondering how to access the dist, point and scores per object that is given by the non_maximum_suppression_inds function per object ? and how does it works ?

The idea behind it to try to use a reinforcment learning approach using agents that can try to maximize the IOU and F1 by changing several parameters for each agent including the distance and the probability for the each object.

How do I get the distance and probability with the metrics per object in the image (are these dist, and point ) or it is something else?
I would like to get these during training for each object at each epoch, so the agents can maximize the metrics, update the segmented objects, and go to the next iteration.

It is conceptual at this point but I am starting to work on it, any inputs, feedback or directions will help :) .
That will help a lot to design the RL approach,
Thanks a lot,

[maweigert]

Hi,

If you want the vertex coordinates and probability for each retrieved object, they are provided in the additional prediction outputs:

y, details = model.predict_instances(normalize(x))

print(details['prob'].shape)
print(details['coord'].shape)

Is that what you want?

[Nal44]

Hi Martin,

yes ! thanks a lot
Basically , the RL algorythm will start after the predictions are made at each epoch, calculate the F1 , or else for each object, take actions to change the masks, replace the masks and go to the next iteration.

So I need the y and the x lists (or tuples ) of all objects, does the x works the same ?
such as :

x, details = model.normalize(x)

print(details['prob'].shape)
print(details['coord'].shape)

Once I get (understand) that , it should be straight forward :) .

Thanks
antho

[Nal44]

Hi ,

I am able to change the coords and points with the agent , now I want to re-calculate the probs for each new object.

How can I feed the updated coords/points values to re-calculate the probs?
Are the probs calculated by the NMS function ? or how the probs are calculated ?

Perhaps I could share in PM my code ?, any inputs will be greatly appreciated :)

Thanks a lot,
antho

[uschmidt83]

Hi @Nal44, sorry for the late reply.

I'm sorry, but I don't understand what you're trying to do.

I am able to change the coords and points with the agent , now I want to re-calculate the probs for each new object.

Have you read the paper or looked at the poster?

The object probabilities and radial directions are predicted by the CNN, which are then used to create a set of object candidates, which the NMS reduces to the final set of predicted objects.

[Nal44]

Hi,

I' ll send you a PM on your email address with the file I am working on , that will be easier to explain and show the logic and the code.

Thanks a lot,
antho

[Nal44]

Hi both,

Pm sent with the corresponding files (2 emails ).
Please let me know :)
thanks a lot,
antho

[Nal44]

Hi,
Just checking did you receive my emails ?
thanks a lot,
Antho