Weird prototxt layer ...

[MyVanitar]

Hello,

where did you got these .prototxt and .caffemodel files?

for example look at this layer at the FCN-8s-OBG-4s.

layer {
  name: "obg-fcn-fuse"
  type: "Eltwise"
  bottom: "mask_projection_no_b"
  bottom: "score_fcn"
  top: "score"
  eltwise_param {
    operation: PROD
  }
}

where this top: "score" refers to? there is not layer in the name of score in the whole text. I think the problem of lower accuracy might come from this or possible similar problems inside prototxt files.

[twtygqyy]

@VanitarNordic Hi, top: "score" is only the name of output layer with inputs of two bottom layers , you can use any name you want.

[MyVanitar]

@twtygqyy

Did you download these from somewhere or you have designed them yourself from paper?

[twtygqyy]

@VanitarNordic This code in this repository was implemented by my understanding according to the first version of obg_fcn paper. The paper and algorithm were updated for ACCV's publication.

[MyVanitar]

@twtygqyy

Good man, it is difficult to do, specially when paper authors give no feedback! a simple mistake or parameter change can lead to a totally different results, better or worse.

May I ask you where can I learn about layers practically and caffe structure to design or implement from paper and be professional like you? sources which explained these good.

[twtygqyy]

@VanitarNordic
https://github.com/kjw0612/awesome-deep-vision contains almost all related paper and material for learning.
I would recommend http://cs231n.stanford.edu/ as a start point.

[MyVanitar]

@twtygqyy

Thank you.

I suggest you to consider this paper instead of BGN. They achieved very good accuracy by modifying the FCN-8s. I asked the authors but they did not handle their trained model or prototxt files but I think it is easy for you modify the existing FCN-8s with their instruction.

https://arxiv.org/pdf/1611.08986.pdf

[twtygqyy]

@VanitarNordic Thanks for the suggestion, I know this paper but didn't try to implement it. I will take a deep look.

[MyVanitar]

@twtygqyy

They have skipped some connections, if I understand correctly they have skipped the classification layers.

[MyVanitar]

@twtygqyy
Hi man, have you reached to any result with that paper?

[twtygqyy]

@VanitarNordic sry for the late reply, was in a long vacation. I've heard that one of my colleagues implemented the improved FCN, haven't tried by myself.

[MyVanitar]

@twtygqyy

So please ask him for the materials, because the implementation is more straightforward than OBG-FCN, because they have improved the existing FCN and named it iFCN-Resnet and reached very good results.