Convolution over only one axis of a multidimensional array (running mean over the frames of a video)

Question

If I have a stack of 1D arrays, it is easy to get a running mean over the first axis:

import numpy as np
from scipy.ndimage import convolve1d
arr = np.random.random(size=(5000,10)) # a stack of 5000 1D arrays, each of length 10
running_mean = convolve1d(arr,np.ones(30)/30,axis=0) # replace each array by an average over 30 of them

However if I have a stack of 3D arrays, the obvious solutions become extremely slow

import numpy as np
arr = np.random.random(size=(5000,250,250,3)) # an rgb video with 5000 images, resolution 250x250
running_mean = np.array([arr[i:i+30].mean(0) for i in range(len(arr)-30)])

Is there a vectorized method in scipy, numpy, or opencv to convolve a kernel with a stack of arrays over only one axis?

Answer 1

You can still use convolve1d for your 3D array and just set the parameter axis accordingly.

Side note: Your method for the 3D array does not handle the case, when you reach the boundaries of the input array. So your output shape will be (4970, 250, 250, 3).