>512 masks cv2 limitation on resize

[fanweiya]

yolov5/utils/segment/general.py

Line 91 in 7398d2d

masks = cv2.resize(masks, (im0_shape[1], im0_shape[0]))

System information (version):

     opencv-python == 4.1.0.25

Detailed description:

     cv2 resize channel limited
     This bug is triggered when the number of masks exceeds 512, and the skimage resize function is recommended.  

refer to:opencv/opencv#14770

[glenn-jocher]

@fanweiya thanks for the bug report! Can you supply a command to reproduce this? Do you know of a suitable workaround for handling >512 predictions?

[fanweiya]

I used Yolov5 to do the instance segmentation of the cell image, which had a lot of cells, more than 512, and then prompted the errorwith gufunc signature (n?,k),(k,m?)->(n?,m?) (size 564 is different from 1), i found the mistake by checking it out.
I used the following solution to replace it,and it worked.

import skimage
masks = (skimage.transform.resize(masks, (im0_shape[0], im0_shape[1], masks.shape[-1]))*255).astype(int)

yolov5/utils/segment/general.py

Line 91 in 7398d2d

masks = cv2.resize(masks, (im0_shape[1], im0_shape[0]))

[glenn-jocher]

@fanweiya wow, ok thanks for the replacement code. I'll create a special if-else here to handle these edge cases.

[glenn-jocher]

@fanweiya I'm a little confused. I ran python segment/predict.py, and got a normal image of shape (w,h,3), no matter how many objects were in the image on L91 of utils/segment/general.py there were always 3 channels in the image, no more. Are you sure this is the right location where cv2.resize was causing an error?

[fanweiya]

When im_gpu=None，the error will be reported

yolov5/segment/predict.py

Line 168 in 350e8eb

annotator.masks(masks,

yolov5/utils/plots.py

Line 137 in c8e5230

masks = scale_image(masks.shape[:2], masks, self.im.shape)

@fanweiya I'm a little confused. I ran python segment/predict.py, and got a normal image of shape (w,h,3), no matter how many objects were in the image on L91 of utils/segment/general.py there were always 3 channels in the image, no more. Are you sure this is the right location where cv2.resize was causing an error?

[glenn-jocher]

@fanweiya Oh it only happens with --retina?