CODEWITHSUNDEEP
pythonimage-processingnumpyfeature-detectionscikit-image
utdemir
utdemir

Reputation: 27236

Fixed-length HOG descriptor for variable sized images

I'm working on some object detection code, however my objects don't have a fixed size, so;

skimage.feature.hog(obj)

doesn't give me equal length vectors(since it uses fixed sized cells), and therefore I can't use learning algorithms on them.

So, I tried dynamically assigning HOG feature length:

from __future__ import division

def describe_object(obj, div=8):
    width, height = obj.shape
    f = skimage.feature.hog(obj, normalise=True,
                            pixels_per_cell=(height//div, width//div))
    return f

But, now it mostly gives 2916 sized vectors, but sometimes it gives longer vectors (like 3402 elements long) too.

I believe this happens when some specific ratio between bin size and object's shape, but don't know why exactly.

Can you help me?

Upvotes: 0

Views: 938

Answers (1)

bjoernz
bjoernz

Reputation: 3852

You could scale the images to a fixed size, before calculating the HOG features.

Upvotes: 3

Related Questions