Detect model return position y is negative

Same as title
i trained my model using all same size data (1920x1080). When create model i set padding as 1920x1080 because i dont want make padding.

Every work fined til i crop image before push it data to GPU. Detect() return position of box have negative value

looked like still have padding on image when Detect() was called, how can i prevent this padding or return negative value ?