How to change nms and confidence threshold in python wrapper? about yolo3-4-py

[VitaliKaiser]

OK I found the answer myself in the code.
This are the function parameters of detect:
def detect(self, Image image, float thresh=.5, float hier_thresh=.5, float nms=.45):

[madhawav]

Hi,
You can filter results using the "score" value (2nd parameter of tuple) provided in results of net.detect() function.

For example, look at the if score > 0.5: line of below sample code.

            dark_frame = Image(frame)
            results = net.detect(dark_frame)
            del dark_frame

            for cat, score, bounds in results:
                if score > 0.5:
                    x, y, w, h = bounds
                    cv2.rectangle(frame, (int(x-w/2),int(y-h/2)),(int(x+w/2),int(y+h/2)),(255,0,0))

Did you try this approach?

[arasharchor]

@madhawav Oh man I totally missed that score. Thanks a lot for your reply.

[arasharchor]

But a small point, the detections are not identical when I use original c++ wrapper. I do not know why

[madhawav]

@smajida Can you send me some more details? Like maybe screen shots on certain scenarios where you noticed a difference?

If you can send, it would be of great help!

[arasharchor]

@madhawav sure. Right now. I do not have my laptop with me. I will send it to you soon. BTW, are you interested in an internship in German Aerospace Center(DLR)?

[madhawav]

@smajida Thank you for the offer! Where can I find more information?

Also, how about opportunities to do higher studies? I recently completed my Bachelors Degree in Computer Science. Right now I am looking for opportunities to do higher studies.

[VitaliKaiser]

Hi,

but how would I lower the minimum threshold? If I get it right, after calling

results = net.detect(dark_frame)

I only get the results above the default confidence threshold of 0.25. I would not get results with a lower threshold. In my case I would like to get, e.g., every result above 0.01. How would I manage that?

Best,

Vitali

[sushant1727]

Is there a way to print the confidence level?

Update: I got it using
cv2.putText(img, str(score), (int(x), int(y)), cv2.FONT_HERSHEY_COMPLEX, 1,(0,255,0),2)

[sushant1727]

@VitaliKaiser @madhawav Where can I find the parameters of net.detect() function & pydarknet files where it imports Detector, Image.
(I installed using pip yolo gpu version.)

NVM, I got it in file pydarknet.pyx

[madhawav]

@smajida Any updates on the screenshots (on scenarios where you noticed detection differences) ?