Spatial regularizer

[vadimkantorov]

Hi Hakan,

Thanks for putting the code online, it's really a nice and concise code base! I've got two questions concerning the spatial regularizer:

1) As far as I can see in https://github.com/hbilen/WSDDN/blob/master/layers/LossTopBoxSmoothProb.m#L69, you are also putting the probs values under the square opposite to the formulation in subsection 3.4 of the paper. Also the gradient computation in https://github.com/hbilen/WSDDN/blob/master/layers/LossTopBoxSmoothProb.m#L85 doesn't square the probs value that doesn't correspond to the forward pass code. Am I missing something or is it a bug?

2) Do I understand correctly that you differentiate the regularizer only wrt fc7 outputs?

Thanks in advance!

[hbilen]

I have updated the equation in the paper. Thanks to Vadim Kantorov and Siniša Šegvić. Now the implementation should be consistent with the paper.

[senthilps8]

Sorry for reviving this. Can you point me to the new version of the paper? Correct me if I'm wrong. From what I understand, this and this version still have the issue that @vadimkantorov pointed out.

[hbilen]

Thanks Senthil, the paper was somehow not updated in the vgg website. I think it should be the right version now. Here is the link:

http://www.robots.ox.ac.uk/~vgg/publications/2016/Bilen16/bilen16.pdf

On 18/12/16 21:20, Senthil Purushwalkam wrote:

Sorry for reviving this. Can you point me to the new version of the paper? Correct me if I'm wrong. From what I understand, this https://www.robots.ox.ac.uk/%7Evgg/publications/2016/Bilen16/bilen16.pdf and this https://arxiv.org/pdf/1511.02853v3.pdf version still have the issue that @vadimkantorov https://github.com/vadimkantorov pointed out.

— You are receiving this because you modified the open/close state. Reply to this email directly, view it on GitHub https://github.com/hbilen/WSDDN/issues/1#issuecomment-267847501, or mute the thread https://github.com/notifications/unsubscribe-auth/AD_sKIBhJB0PMZlUidNH8Yh5oinAzit6ks5rJaOBgaJpZM4I99Zz.

[vadimkantorov]

Still, from what I can see, the forward pass squares the probs while the backward doesn't. Shouldn't they be consistent on this?

[senthilps8]

Yes, I think the backward prop uses the previous formulation of the loss #4

[hbilen]

Thanks, you are right. I guess if we square probs term, it is gonna be correct, isn't it? This might require some change in the scaling term. I will try to run an experiment with that.

[vadimkantorov]

Yep, seems squaring probs in backward pass would make it correct

[css1995]

@vadimkantorov Sorry, could you show me the code after squaring probs in backward pass?