Troubleshooting (-15:Bad number of channels) for face landmarks detection

Python
1

Hello I am doing a project using LBF landmark detector using the sample landmark_demo.py as a starting point.

the detector is able to successfully detect faces on the 512 x 512 image of lena as well as a custom 640 x 477 image. however when I tried to run the detector on a 320 x 240 image I get the below error.

 img = array([[[171, 169, 163],
        [170, 170, 164],
        [170, 170, 164],
        ...,
        [160, 160, 161],
     ...[151, 152, 148],
        ...,
        [142, 143, 167],
        [143, 144, 168],
        [144, 144, 164]]], dtype=uint8)  

     def get_face_landmarks(img):
        faces = FACE_DETECTOR.detectMultiScale(img,
                                               scaleFactor=1.3,
                                               minNeighbors=4,
                                               minSize=(30, 30),
                                               flags=cv2.CASCADE_SCALE_IMAGE
                                               )
       _, landmarks = FACE_LANDMARK_DETECTOR.fit(img, faces=faces)
E     cv2.error: OpenCV(4.8.0) D:\a\opencv-python\opencv-python\opencv\modules\core\src\matrix.cpp:1193: error: (-15:Bad number of channels) The total width is not divisible by the new number of channels in function 'cv::Mat::reshape'

I know it has something to do with my image resolution but I am unsure how to diagnose it further.

My questions are:

  • what is causing this exception?
  • is there a parameter I can set in the facemark.fit function?
  • Is there any way to resolve this issue without having to resize the image? (I want to have low latency and reshaping the image would probably add computational delays.)
2

Turned out the problem wasn’t with the image size but whether or not the image contained a face. If no faces are detected faces is an empty tuple and facemark.fit does not handle that being empty.

I verified this by modifying the landmark_demo.py script with a non-face image and was able to reproduce the issue.

I ended up fixing this by putting a conditional before the fit in case no faces are detected.

1 Like
3

nice catch !
i’d even argue, that you found a bug in the sample,
(it MUST check, if there ARE any faces !)

time to make a pr and change this !

4

Hi berak, I’ve made a pr here. I’m pretty new to contributing so please let me know if there’s anything missing. Thank you.

1 Like
5

looks great ! thank you :wink: