LinAlgError due to Singular Matrix

[Make42]

With some images the algorithm

import saliency_triple as psal
psal.get_saliency_mbd(filename).astype('uint8')

throws an LinAlgError in

cov_left = np.linalg.inv(cov_left)
cov_right = np.linalg.inv(cov_right)

cov_top = np.linalg.inv(cov_top)
cov_bottom = np.linalg.inv(cov_bottom)

since one of cov_... can be singular. Using np.linalg.pinv does not work - the saliency map becomes solely black. I do not understand the algorithm good enough to improve one that. Could you please?

[yhenon]

Hi, Could you share an image that causes this issue? Thanks.

[Make42]

Yes. These images caused the issue:

Please try both.

[Make42]

No progress?

[yhenon]

Hi, I haven't been able to replicate this issue. Using your image as test1.jpg

import pyimgsaliency as psal
import cv2

filename = 'test1.jpg'

mbd = psal.get_saliency_mbd(filename).astype('uint8')
rbd = psal.get_saliency_rbd(filename).astype('uint8')

cv2.imwrite('mbd.png',mbd)
cv2.imwrite('rbd.png',rbd)

Succesfully saves the images to disk (note that these algos are not designed for medical images, results may not be the best).

Can you please check that your packages are up to date (numpy, scipy and sklearn)? Can you also see if the source image is in a stange format (this does happen with medical images)? Input should be 8-bit RGB images.