lucasb-eyer
commented
7 years ago I have never done this, but the whole point (pun intended) of denseCRF is to densely connect all points with all.
I have just spent some time going through the code and the paper, and I don't think there's a way to do this. If you re-read sections 3 and 4 of the paper, you'll see that nowhere is there a term where you could specify individual edge weights. Possibly, you could hack around this in some way by defining some specific kind of features and full kernel precision-matrix. That'd be a horrible hack and I'm not sure it'd work at all, though.
What you want sounds like a more specific type of CRF, not a DenseCRF.
Hi, thank you very much for code. But I have some questions when I use this code.
My task is to cluster some points, not image. Here some points are connected but some are not. Say, I have 100 points. So I build a crf model by code below(in your demo).
d = dcrf.DenseCRF(100, 5) # npoints, nlabels
feats = np.array(...) # Get the pairwise features from somewhere. print(feats.shape) # -> (7, 100) = (feature dimensionality, npoints) print(feats.dtype) # -> dtype('float32')
dcrf.addPairwiseEnergy(feats)
My question is, here how could I set the point connectivity？ For example, point 1 is connected with point 2, but not point 3. Can I set the edge? If I can't, does this mean that all the points are connected?
Looking forward to your reply. Thank you!