CODEWITHSUNDEEP
imagemachine-learningcomputer-visionprobability
kcc__
kcc__

Reputation: 1648

Updating the parameter of Conditional Random Fields without re-training

I am working on a project where we are developing and application for segmentation of foreground object on a mobile device.

Because the segmentation can never be perfect, we are allowing the user to manually adjust the true region after the initial segmentation. The initial segmentation is achieved using pre-trained Conditional Random Field(CRF).

What we want to do further is that once the user adjust the foreground region after the initial segmentation, we want to update the weight parameter of CRF without retraining. However we are not sure if we can achieve this? Is there any approach to do this? Papers?

Upvotes: 0

Views: 124

Answers (1)

Bharat
Bharat

Reputation: 2179

If you are using something like a graph cuts based model, all you would need is to have some hard seeds which would correspond to the user defined regions. For these regions, you can set the unary potentials to infinite and run your CRF inference procedure again. This will not require changing the weights or re-training the CRF.

Upvotes: 1

Related Questions