Wrong output when confidence score is relatively low

[Patinetilla]

Dear Eric,

First of all, many thanks for such great repo (together with QReader).

By chance, I have noticed an erratic behavior. When the confidence score is relatively low the values of 'wh' are wrong and return a number close to the width and height of the picture. For confidential reasons, I cannot show you an example but I have plotted the score and the w and the h of the detect_and_decode method of 30 images:

Do you have any idea why this is happening? Am I missing something?

Many thanks!

[hila-sony]

I noticed something similar too, even on high confidence detections, sometimes the coordinate values are way above the image height\width (specifically it was padded_quad_xy, that is later used to crop the detection before the perspective correction). In my case it caused a bug as the warping inside the decode method relies on these coordinates, and it consumed too much memory. I can provide the test image privately.