search

sadeepj / crfasrnn_keras

CRF-RNN Keras/Tensorflow version
http://crfasrnn.torr.vision
MIT License
603 stars 169 forks source link

Alternative repository #34

Closed MiguelMonteiro closed 5 years ago

MiguelMonteiro commented 6 years ago

Just wanted to let you know that I developed an alternative repo which is more flexible in the shape and number of channels of the images (any 2D, 3D multi-channel image should work) and that has both a CPU and GPU kernel. Please check the link if you are interested in more info. https://github.com/MiguelMonteiro/CRFasRNNLayer

iperov commented 6 years ago

so is there no way to port permutohedral_lattice to native tensorflow ?

MiguelMonteiro commented 6 years ago

That is the second repository I built. It is linked in this one. It's not so much a port but more of complete overhaul of the code to optimize it and remove bugs I found along the way. Here you go:

https://github.com/MiguelMonteiro/permutohedral_lattice

glhfgg1024 commented 6 years ago

Hi @MiguelMonteiro , great work! Thanks a lot for your sharing!

MiguelMonteiro commented 6 years ago

@glhfgg1024 Btw I did try this for 3D medical imaging segmentation and couldn't get a statistically significant performance increase. Let me know if you observe something different.

glhfgg1024 commented 6 years ago

@MiguelMonteiro , yes, in my case, the improvement on dice was also limited (1%-2%), and my MICCAI submission has been rejected. I agree with you that the expert contours in medical images have less defined edges as those in natural images. For example, for lung CT tumor segmentation, the expert contours include many voxels with -100~-800 HU values. And when using the CRF, the module still cannot correct the mis-classified voxels and may even lead to worse segmentation.

iperov commented 6 years ago

can permutohedral_lattice be replaced by something implemented by native tf ?

MiguelMonteiro commented 6 years ago

Well it's just a bilateral filter or a gaussian filter. I am sure you could implement a brute force implementation in native tensorflow if you really wanted to. However, I think it is best to keep questions regarding this topic on the permutohedral_lattice 's page.