Bounding box out of the image size using CascadeObjectDetector

I am using CascadeObjectDetector for traffic sign detection with the GTSDB public dataset.

I trained the detector, but now, when I pass an image of 800x1360x3 uint8, the detector returns me a bbox like [979 828 386 187] which is out of the bounds of the original image and seems to be impossible (even the width and the height of the bounding box are suspiciously big). So, when evaluating the performance of the detector the result is a false positive.

Is this the normal behaviour and shall I just discard these the bboxes that are out of the bounds of the images or is this a bug of the detector?