Inconsistent of adjust_contrast and adjust_brightness with PIL lib

The adjust_contrst implementation in Kornia is essentially the adjust_brightness in Torchvision.

In order to reproduce the results from Torchvision, I think we might need to rename the adjust_contrast to adjust_brightness, and rewrite the adjust_contrast. Our current implementation of adjust_contrast is not adjusting contrast at all.

To reproduce:

from PIL import Image
import requests
from io import BytesIO
from torch import allclose
from kornia.color.adjust import (
    adjust_brightness, adjust_contrast, adjust_saturation, adjust_hue)

response = requests.get("https://tinypng.com/images/social/website.jpg")
img = Image.open(BytesIO(response.content))

import torch

import kornia.augmentation.functional as F
from torchvision.transforms import functional as tvF
from torchvision.transforms import transforms

import matplotlib.pyplot as plt
import numpy as np

to_tensor = transforms.ToTensor()
to_pil = transforms.ToPILImage()

def tensor_pre_transform_wrapper(input: torch.Tensor):
    """ A wrapper that tried to reproduce the actual output from:
        - transforms.ToPILImage()
        - transforms.ToTensor()
        For each image, simply (img * 255).int() // 255
    """
    return (input * 255).int().float() / 255

in_tensor = to_tensor(img)
in_pil = img
# factor = 0.5
out_tensor = adjust_contrast(tensor_pre_transform_wrapper(in_tensor), torch.tensor(2))
out_pil = tvF.adjust_brightness(in_pil, 2)

allclose(out_tensor, to_tensor(out_pil))

[shijianjian]

Yes. I will finish the tests for spatial augmentation in this weekend and update the docs accordingly.

[shijianjian]

Closed since we are going to keep those functions in OpenCV math than PIL.