Mtcnn accuracy problem

[SaddamBInSyed]

Thanks for your work.

I have carefully followed your build instructions and when I run the TensorRT mtcnn face detector on my jetson nano, I am getting low accuracy as like below
code detecting my hand as a face (continuously for some seconds ).

please advise.

[jkjung-avt]

I think my TensorRT MTCNN result should be very close to the original Caffe model.

I did not train the original Caffe MTCNN model by myself. If you'd like to improve its recall/precision, please look into how to re-train or fine tune it.

[julayelei]

Hello, Saddam. You can also change values of MINSIZE for faces detected. Sometimes, some objects "look" like a face. By default, it uses 40 minzise in utils/mtcnn. Try changing it to a higher minsize!

[SaddamBInSyed]

@jkjung-avt
thanks for your advice

seems it works fine when I change the minsize from 40 to 100 pixels.

thank you.

[jkjung-avt]

Thanks to @julayelei