detector.detectCustomObjectsFromImages returns negative numbers in box_points detection about imageai HOT 2 CLOSED

Ok, so I think it is an opencv issue. My solution is to just set the negative values to 0 in the box_points array, so that the opencv copy method can do it's job. Not sure if it works in general but on my dataset it is usually the top of the picture that is cropped to hide peoples faces and I guess the pretrained resnet50_coco_best_v2.0.1.h5 will tend to detect people with their faces not cropped. :)

My solution:
In imageai/Detection/init.py line 743 I add:
detection_details = [0 if x < 0 else x for x in detection_details]

from imageai.

I am happy you found a way around the issue. I will investigate the issue further to see if a more solid solution can be provided, else I will adopt your approach of keeping the negative value at 0.

Also, the RetinaNet model (resnet_coco_best_v2.0.1.h5) can detect persons irrespective of the perspective they are viewed from.

from imageai.