how can i implements landmark image layer in caffe about deepalignmentnetwork

[MarekKowalski]

Hi, I never used caffe, so it is hard for me to help with the implementation itself.
I advise that you take the look at the implementation in LandmarkImageLayer.py and try to reproduce it. Since this is fully deterministic the results you will get should be very similar to what this layer produces.

I will be more than happy to answer any specific questions as well.

Marek

[spark123456789]

As I can see, the landmark image layer generate the heatmap for the outupt layer, to enhance the information of predicted position (landmark) from previous stage. It is generally a forward process, this park shuold be easily implemented in caffe, my question is how can i implement the backward process, from heatmap to landmark. Can you give some advise?Thank you.

[MarekKowalski]

Hi, I am not sure by what you mean in "from heatmap to landmark".
The heatmap is used as an input to a convolutional neural network which infers the updates to the landmark locations based on the heatmap and other data, there is no single "heatmap to landmark" procedure.

Marek