Port segnet architecture to caffe/master branch?

[anandthakker]

Since the Caffe main branch now has a batch normalization layer and upsampling via deconvolution layer, it seems to me that it might be feasible to port the SegNet architecture so that it can run off the BVLC/caffe master (using Python layers for the data input layer and per-class-weighted loss).

I wonder if this might be worthwhile, because, while I haven't followed it in detail, it seems that caffe has gotten any number of improvements in the past year that segnet training could benefit from.

@alexgkendall what do you think?

[alexgkendall]

I agree this would be great - aside from what you've pointed out you would also need to port the max unpooling layer too.

[arassadin]

IMHO, integration with Caffe is important for growth since Nvidia is taking steps towards this framework, such as Digits or GIE and SegNet's idea and implementation still original and useful.

[anandthakker]

I agree this would be great - aside from what you've pointed out you would also need to port the max unpooling layer too.

@alexgkendall From a naive read of the prototxt file, I don't see these as a layer type -- could you point me to an example of one of these layers?

[jorgyz]

@anandthakker I believe @alexgkendall is referring to the "upsample" layer, which is the same thing as unpooling. There is a discussion on the Caffe User's Group about incorporating an unpooling layer into the master branch (see https://groups.google.com/forum/#!searchin/caffe-users/unpooling%7Csort:relevance/caffe-users/rB5jSwqrwDY/2z-z67FXAQAJ) but I don't know whether there has been any traction on that.

[ferrouswheel]

One problem might be that SegNet's license doesn't match Caffe's license, the latter being fine with commercial use.

SegNet says:

"This extension to the Caffe library is released under a creative commons license which allows for personal and research use only. For a commercial license please contact the authors. You can view a license summary here: http://creativecommons.org/licenses/by-nc/4.0/"

Of course @alexgkendall could dual-license (retaining copyright of his code) and allow the Caffe version to be used more permissively, but without that agreement I doubt Caffe would accept any pull request with SegNet code. The other option is that someone re-implements the missing bits.