Detection output boxes is out-of-bound

[pasin-k]

Thank you for the resource you published.
So I tested the data with some of my private fisheye data.
I used the pretrain-weight to see how it looks initially and I noticed that for detection result, there are significant amount of boxes that is out of bound (around left-bottom corner) that has very high score and almost none elsewhere. It could be that it is just pre-trained model which not giving best result but I want to confirm if it is not some misconfiguration bugs.

For some information, my dataset is slightly in different view with the provided dataset and the resolution is different but it is a car-attached camera view so I suspect there might be something I need to adjust other than cropped_coords which I adjust to my image but it does not help at all.

Any idea?