Keras: R-CNN ROI pooling layer

Created on 22 Apr 2016 · 11Comments · Source: keras-team/keras

Someone has a theano port of the layer that was custom implemented for Ross Girschick's fast/faster RCNN models: https://github.com/ddtm/theano-roi-pooling

I thought I'd leave this here in case someone wants to pick that up and run with it to integrate it into Keras. I can't work on it myself until June at the earliest.

stale

Source

braingineer

Most helpful comment

I am working on implementations here

yhenon on 10 Jan 2017

👍6 🎉3

All 11 comments

@braingineer I'm new to keras, I am interested in implementing/porting ROI-pooling too, but I'm using Tensorflow Back-end for the rest of the work.
Do you plan to do it on keras as Tensorflow backend? or in a way that could benefit both backends (if there is a way to hack it)

Irtza on 2 May 2016

Hi @Irtza , I wasn't planning on it because I don't really use tensorflow (as least, not in the near future). But, in general, I wasn't planning on getting to this until at least after the first few weeks in June. There is way too many other things first.

If you want to port it, you can always ask the mailing group or slack rooms for advice if you get stuck.

braingineer on 2 May 2016

@braingineer I have some time on my hand in June. but the tensorflow layers could be really tricky to implement, I'm not so comfortable with advanced C++ that needs to be written to implement r-cnn through tensorflow. however porting the theano layer to keras sounds like a do-able task, and could be of help to me. I need some guidance.
where is the a mailing group or slack room for keras?

Irtza on 5 May 2016

Hi @Irtza I'm interested in doing ROI pooling too. I'd love to help you out in this.