vision.ImageSegmenter fails when image array is F-continguous

[lillekemiker]

When using numpy in python, the ordering of array elements in memory is handled under the hood and usually the user never has to worry about this. For instance, when I read an image file from disk parse it as a numpy array with image = np.asarray(Image.open('image.jpg')), it is loaded in C-contiguous format. If I decide to crop the image by slicing (crop = image[20:-20, 30:-3]) I will get an F-contiguous array as a result.

If I try to run segmentation inference using the vision.ImageSegmenter.segment() method with a C-contiguous array as input, it works fine. If the input image array is F-contiguous, the model happily returns garbage.

Versions: numpy: 1.23.1 python 3.9.10 tflife-support: 0.4.2 tensorflow: 2.8.3

Suggested fix: Add the following to the inference method:

    if not image_array.data.c_contiguous:
        image_array = np.ascontiguousarray(image_array)

[lu-wang-g]

Thanks for the suggestions!

We'll add this update to MediaPipe Tasks, which will have a preview release soon. Before that, please work around this issue by passing in C-contiguous array specifically to the Task Library API.