Masks with more than one layer about kaggle-dsbowl-2018-dataset-fixes HOT 7 OPEN

Hi @John1231983 , you can apply some small changes to your code simply defining a custom reading method and passing it to the ImageCollection class constructor. I've fixed it this way:

def _imread_mask(f):
    mask = skimage.io.imread(f)
    if len(mask.shape) > 2:
        mask = mask[:,:,0]
    return mask

...   

masks = skimage.io \
    .ImageCollection(masks_path, load_func = _imread_mask) \
    .concatenate()

from kaggle-dsbowl-2018-dataset-fixes.

I'm using imread and not imread_collection from skimage. That's probably why I don't have this problem. Two solutions:

• Alter the images directly (and do a pull request :) ).
• Don't use concatenate() and perform concatenation yourself after using the trick above

from kaggle-dsbowl-2018-dataset-fixes.

Hi @ebouteillon You can take the top layer of the mask (mask[:,:,0]), layers (0,1,2) has max pixel set for the mask.

from kaggle-dsbowl-2018-dataset-fixes.

Hi @PrinceRK , It is already what I am doing ;)
It is just a notice that these masks are not consistent with others (and the originals by the way).
Regards

from kaggle-dsbowl-2018-dataset-fixes.

@ebouteillon : How do you solve it? I am run the code to classify the dataset into 2 cluster but I got the error about masks mismatch dimension of these above ids. This is what I am doing

# Get image width, height and count masks available.
def read_image_labels(image_id, space="rgb"):
    image = read_image(image_id, space=space)
    mask_file = STAGE1_TRAIN_MASK_PATTERN.format(image_id)
    print (mask_file)
    
    masks = skimage.io.imread_collection(mask_file).concatenate()
    height, width, _ = image.shape
    num_masks = masks.shape[0]
    labels = np.zeros((height, width), np.uint16)
    for index in range(0, num_masks):
        labels[masks[index] > 0] = 255  # index + 1
    return image, labels, num_masks

from kaggle-dsbowl-2018-dataset-fixes.

Instead of using masks[index] directly, I'm using an intermediate object. Let's call it m, then it would give you something like:

m = masks[index]
if(len(m.shape) > 2):
    m = m[..., 0]

You may have to use np.squeeze too if the extra dimension annoys you.

from kaggle-dsbowl-2018-dataset-fixes.

Thanks for your solution. But the error come from the line

 masks = skimage.io.imread_collection(mask_file).concatenate()

from kaggle-dsbowl-2018-dataset-fixes.