Threshold setting or control for Detection score in Object detection[DNN and MMOD].

[harshkn]

Hello, I have trained an object detector based on DNN MMOD loss layer in DLIB to detect "A" in an image. For high res images, I can detect "A" in an image. But in case of low resolution images, I am not able to detect. In order to understand why, I tried to visualise the detector output scores and I get the below result.

I can see that I get good score for some pyramid scale(bottom left in the pyramid), but the output of the detector does not return any rectangles. I am guessing, there is a threshold that is set and the score is not above the threshold. So my question is, is my guess right, if yes, do I need to look into loss.h file or any other file ? (please let me know) and if not, how do I go about getting the mmod_rect as result on "A".

[davisking]

You could certainly change the detection threshold. You will probably get false alarms as well though. But in any case, you can easily adjust the threshold by calling net.process(). See the documentation http://dlib.net/dlib/dnn/core_abstract.h.html#add_loss_layer

[harshkn]

Thank you very much @davisking for the help. This is what I was looking for.