Computing all the Haar-like features using scikit-image

Question

How can we compute all the Haar-like features of all types using scikit-image function haar_like_feature? This is what I have tried (a simple example for computing all the features of type 2x):

from skimage.feature import haar_like_feature
from skimage.transform import integral_image

img = np.array([[1,  2],
                [1,  3]])

ii = integral_image(img)

features = haar_like_feature(ii, 0, 0, ii.shape[1], ii.shape[0], 'type-2-x')

print(features)
[1, 2]

However I would expect to get [1, 2, 3], because we should also consider a rectangle feature covering the whole image resulting in the feature value (2 + 3) - (1 + 1) = 3.

I also checked the Viola-Jones paper and they have the following number of features:

type-2-x: 43200
type-2-y: 43200
type-3-x: 27600
type-3-y: 27600
type-4: 20736

total: 162336

(source: An Analysis of the Viola-Jones Face Detection Algorithm)

However the number of features produced by skimage.feature.haar_like_feature is different:

img = np.random.randint(0, 256, (24,24))
ii = integral_image(img)

total = 0
for feature_type in ['type-2-x', 'type-2-y', 'type-3-x', 'type-3-y', 'type-4']:
    features = haar_like_feature(ii, 0, 0, ii.shape[1], ii.shape[0], feature_type)
    print(f"{feature_type}: {len(features)}")
    total += len(features)
print("\ntotal:", total)

:

type-2-x: 43056
type-2-y: 43056
type-3-x: 27508
type-3-y: 27508
type-4: 20736

total: 161864

So it seems that there are 472 features missing in this computation. Am I doing it wrong? What parameters should I pass to the function haar_like_feature() to get all the features?

Update: there seems to be a bug in the implementation of the function haar_like_feature https://github.com/scikit-image/scikit-image/issues/4818

Answer 1

for here,just wanna share:

from here : img = np.array([[1, 2], [1, 3]])

to : img = np.random.randint(0, 256, (12,12))

try to increase the input size of np.array() or image size..to for example 12x12.