there is a bug in inference code
change
scale_img=cv2.resize(img,((int(img.shape[0]*scale)),(int(img.shape[1]*scale))))
to
scale_img=cv2.resize(img,((int(img.shape[1]*scale)),(int(img.shape[0]*scale))))
to fix this bug
you can find more details in my blog below.
CSDN :https://blog.csdn.net/Rrui7739/article/details/81261543