Rotate a batch of images using `scipy.ndimage.interpolation.rotate`

Question

Let's suppose that X is a numpy.ndarray tensor that holds a set of color images. For instance, X's shape is (100, 3, 32, 32); i.e., we have 100 32x32 rgb images (such as those provided by the CIFAR dataset).

I want to rotate the images in X and I found out that this could be done using the scipy.ndimage.interpolation.rotate function of scipy.

However, I can do it correctly only for a single image and a single color channel each time, while I would like to apply the function for all images and all color channels at once.

What I tried and it works so far is the following:

import scipy.ndimage.interpolation

image = X[0,0,:,:]
rotated_image = scipy.ndimage.interpolation.rotate(input=image, angle=45, reshape=False)

but what I want to do, which could look like this

X_rot = scipy.ndimage.interpolation.rotate(input=X, angle=45, reshape=False)

does not work.

Is there any way to apply scipy.ndimage.interpolation.rotate for a batch of images like the numpy.ndarray tensor X?

In any case, is there any better way to rotate a batch of images (stored in a numpy.ndarray) efficiently?

Answer 1

If you want a one-liner,

import scipy.ndimage.interpolation as intrp

rotated_imgs = [intrp.rotate(input=img, angle=45, reshape=False) for img in X]

assumed your X has the more standard shape (n_images, w, h, color_channels).

If you want to get your hands on matrices, I suggest you to generate your rotation matrix with cv2.getRotationMatrix2D, replicate/stack it along the first axis, and then apply it to your X, but it can get complicated without cv2.warpAffine, which takes a single image as input.