Colorspace reversal hiccup

Question

I often deal with rgb->bgr issues due to PIL vs. OpenCV loading images differently. Usually I can just do

orig = orig[...,::-1]

but I have run into a strange hiccup down the line with that (i'm using OpenCV 3.2) ; a later function for adding some boxes to the image hits

TypeError: Layout of the output array img is incompatible with cv::Mat (step[ndims-1] != elemsize or step[1] != elemsize*nchannels)

The various approaches with everything except the one that works commented, shown. The image after rgb->bgr looks fine, has correct dims. The deepcopy was due paranoia that the problem was caused by some change to image_np.

  image_np = load_image_into_numpy_array(image)
  bgr_img = copy.deepcopy(image_np)
  print('origshape {}'.format(bgr_img.shape))
  #bgr_img  = bgr_img[...,::-1] #fails
  #bgr_img  = bgr_img[...,[2,1,0]])   #fails
  #bgr_img  = np.array(bgr_img[...,[2,1,0]])   #fails
  bgr_img = cv2.cvtColor(bgr_img,cv2.COLOR_RGB2BGR) #works
  cv2.imshow('bgr',bgr_img)
  cv2.waitKey(0)
  print('finalshape {}'.format(bgr_img.shape))
      ...
      

The function that fails is

def bb_with_text(img_arr,bb_xywh,text,boxcolor=[50,255,50]):
    text_color=[0,50,255]
    text_bgnd_color=[220,255,180]
    cv2.rectangle(img_arr,(bb_xywh[0],bb_xywh[1]),(bb_xywh[0]+bb_xywh[2],bb_xywh[1]+bb_xywh[3]),color=boxcolor,thickness=2)
    img_arr[bb_xywh[1]:bb_xywh[1]+20,bb_xywh[0]:bb_xywh[0]+ bb_xywh[2]]=img_arr[bb_xywh[1]:bb_xywh[1]+20,bb_xywh[0]:bb_xywh[0]+bb_xywh[2]]/2)+np.array(text_bgnd_color)/2
    cv2.putText(img_arr,text,(bb_xywh[0]+5,bb_xywh[1]+20),cv2.FONT_HERSHEY_PLAIN, 1, text_color)
    return img_arr

with

-> 2063     cv2.rectangle(img_arr,(bb_xywh[0],bb_xywh[1]),(bb_xywh[0]+bb_xywh[2],bb_xywh[1]+bb_xywh[3]),color=boxcolor,thickness=2)
   2064     img_arr[bb_xywh[1]:bb_xywh[1]+20,bb_xywh[0]:bb_xywh[0]+bb_xywh[2]]=(img_arr[bb_xywh[1]:bb_xywh[1]+20,bb_xywh[0]:bb_xywh[0]+bb_xywh[2]]/2)+np.array(text_bgnd_color)/2
   2065     cv2.putText(img_arr,text,(bb_xywh[0]+5,bb_xywh[1]+20),cv2.FONT_HERSHEY_PLAIN, 1, text_color)

TypeError: Layout of the output array img is incompatible with cv::Mat (step[ndims-1] != elemsize or step[1] != elemsize*nchannels)

Its not a showtopper as the cv2 convert does the trick but it is deeply troubling