Log polar transformation in torchvision.transforms

🚀 The feature

Log polar transformation for images, which maps an image from the 2D Cartesian coordinate system to the log-polar system. This transformation each pixel [x, y] to the (ρ, θ) coordinate system, where ρ is the log of the distance from the center and θ is the angle.

Motivation, pitch

I’ve been researching ways of making one of my CNNs rotation invariant and happened to find this paper. I looked all over the docs for a similar transformation but couldn’t find any, so I was wondering if this transformation could be included in the torchvision.transforms module in order to make it easier to do this type of preprocessing on images.

Alternatives

No response

Additional context

I already have a minimal implementation for my personal use that is based on existing transformations in the package. If you think this feature will be useful, I could polish it up and open a pull request to merge it.

cc @vfdev-5 @datumbox

[vfdev-5]

@aftoul meanwhile you can use the following code as a custom transformation for you use-case. This implementation uses OpenCV to warp the image:

Code

import numpy as np
import cv2
from PIL import Image


def to_log_polar(img, max_radius=32.0):    
    assert isinstance(img, Image.Image)
    
    dsize = img.size
    center = [s // 2 for s in img.size]
    flags = cv2.WARP_POLAR_LOG
    out = cv2.warpPolar(
        np.asarray(img), dsize=dsize, center=center, maxRadius=max_radius, flags=flags
    )
    return Image.fromarray(out)


def from_log_polar(polar_img, max_radius=32.0):
    assert isinstance(img, Image.Image)

    dsize = polar_img.size
    center = [s // 2 for s in polar_img.size]
    flags = cv2.WARP_POLAR_LOG | cv2.WARP_INVERSE_MAP
    out = cv2.warpPolar(
        np.asarray(polar_img), dsize=dsize, center=center, maxRadius=max_radius, flags=flags
    )    
    return Image.fromarray(out)


class ToLogPolar:

    def __call__(self, image):
        return to_log_polar(image)

    
class FromLogPolar:
    
    def __call__(self, image):
        return from_log_polar(image)

Usage

from torchvision.datasets import MNIST
import matplotlib.pyplot as plt

dataset = MNIST("..", download=False)


plt.figure(figsize=(20, 7))
for i in range(4):
    img, target = dataset[i]
    plt.subplot(1, 4, i + 1)
    plt.imshow(img)
    plt.title(f"Target: {target}")  
    

t = ToLogPolar()

plt.figure(figsize=(20, 7))
for i in range(4):
    img, target = dataset[i]
    plt.subplot(1, 4, i + 1)
    plt.imshow(t(img))
    
t_inv = FromLogPolar()

plt.figure(figsize=(20, 7))
for i in range(4):
    img, target = dataset[i]
    plt.subplot(1, 4, i + 1)
    plt.imshow(t_inv(t(img)))

Output

Original:

Log-Polar:

Reconstruction:

HTH

[vfdev-5]

Maybe, it is about providing Log-Polar and its inverse transformations. An image from the paper on MNIST: Image from the paper…