GC-86: Add support for the new Haar classifier definition format in OpenCV

[cesarsouza]

The OpenCV code has been recently updated to support a new Haar cascade definition format. It would be nice if the framework could support this new format as well.

Examples of the two different formats are available at

• https://github.com/Danukeru/FOUCAM/blob/master/haarcascade_frontalface.xml
• https://github.com/Itseez/opencv/blob/master/data/haarcascades/haarcascade_frontalface_alt.xml

In order to add support for the new format, it would be necessary to implement a new HaarCascadeSerializationObject (in https://github.com/accord-net/framework/blob/development/Sources/Accord.Vision/Detection/Haar%20Cascade/HaarCascadeStage.cs ) and, most likely, rewrite the existing XML attributes over the classes methods with the different schema. Compatibility with the older format must be preserved.

[JimBobSquarePants]

I'm glad this is on your radar. I had a quick look at the two different format but couldn't figure out the correlation between elements. Only very few elements like leafValues and left_val, right_val seemed obvious to me.

[hzawary]

Dear Dr,

I found the issue and your advice (support for the new Haar classifier definition format in OpenCV) for support new XML training file formats in Accord. I need to use it, but unfortunately the issue isn't done yet! May it done in closest time?! I also implements a interface tool in my repository for create xml training files by OpenCV facilities by a convenient WPF app.

Best regards, Hashem :)

[JimSEOW]

@cesarsouza @JimBobSquarePants Will this link help? Perhaps someone here has tried?

Any update to this issue ?

[cesarsouza]

Hi @JimSEOW,

Actually this issue is quite low priority in my current to-do list (the first one right now being the TensorFlowSharp integration - #92 and #279) .

Please, may I ask where (in which circumstances) do you actually need to use the new Haar Cascade format? I am not giving that much priority to this issue because I understand the code still works fine with the previous format (which had already plenty of working different classifier definitions at the time).

However, if you have a very interesting use case in mind, I might give it a bit more attention (or at least help with the particular application you are working with).

Regards, Cesar

[JimSEOW]

@cesarsouza The goal is to get a detailed facial landmark compatible to what can be provided by e.g. Kinect. I understand your commitment to TensorflowSharp. Definitely it is of highest priority. I already contacted the author of Dlib PInvoke wrapper (68 landmark features detection) perhaps we can get a .NET friendly wrapper for x86, X64 and ARM