CNN in dlib - mmod_human_face_detector

[ntd1683]

Hello, I'm a student,
I want to learn about how cnn works in dlib when using cnn_face_detection_model_v1 mmod_human_face_detector

Can you help me, or read its processing flow ?

Thank you very much

[arrufat]

Sure, check this article: https://blog.dlib.net/2017/08/
Even though it's for cars, it works the same as the one trained on faces.

[ntd1683]

Is there any way I can read the mod file mmod_human_face_detector.dat

[arrufat]

I don't know what you mean by "read" the file, just use it as in the example:

• C++: http://dlib.net/dnn_mmod_face_detection_ex.cpp.html
• Python: http://dlib.net/cnn_face_detector.py.html

If you want to inspect the weights, instead, you can use dlib::net_to_xml.
Here's a blog post that does that to visualize the learned convolutional filters:

• https://jacobgil.github.io/deeplearning/hallucinating_faces_dlib_pytorch

If you want to actually see the model architecture, you can use dlib::net_to_dot.

[ntd1683]

Thank you very much , This is exactly what I'm looking for

[ntd1683]

Can you tell me where you found this photo? I want to know so I can cite the source in the report

[arrufat]

This image was generated using the dlib::net_to_dot function in the C++ example I mentioned above.

    net_type net;
    deserialize(argv[1]) >> net;  
    net_to_dot(net, "mmod_human_face_detector.dot");

Then:

$ dot -T png < mmod_human_face_detector.dot > mmod_human_face_detector.png

You can also generate a PDF:

$ dot -T pdf < mmod_human_face_detector.dot > mmod_human_face_detector.pdf

You must have Graphviz installed for this to work (the dot executable is part of Graphviz).

Here's the generated file, so you can modify it and re-run the dot command on it: mmod_human_face_detector.dot

For the citation, just cite dlib.

[ntd1683]

thank you very much