[BUG] Hue color jitter on an all gray or all white image returns a black image

[benlansdell]

Describe the bug

Hue jitter, via cucim.core.operations.color.color_jitter, on images with a lot (or all) gray seems to overflow and give either black pixels, or an all-black image.

I noticed this when using augmenting images of planes in, e.g., CIFAR: color_jitter with a non-zero hue parameter would render some sky pixels black. I found this issue only comes up when the hue parameter is non-zero -- other jitter parameters didn't seem to affect the behavior. Further experimenting showed that inputting an image of all white or all one shade of gray resulted in color_jitter outputting an image of all zeros.

More specifically:

import cupy as cp
import cucim.core.operations.color as ccl

x_cupy = cp.ones((3, 64, 64)).astype(cp.uint8)*255
x_jitter = ccl.color_jitter(x_cupy,0,0,0,0.1)

#Outputs all zeros, i.e.
(x_jitter == 0).all() # is True

The same behavior occurs for any gray image, e.g. if I make x_cupy a tensor of all intensity 128.

Perhaps this is some rounding/overflow issue since color_jitter seems to convert all inputs to uint8?

Expected behavior

I would expect the output to not be all black but, if the image is all gray then hue jitter should have no effect -- it should output the same gray image. The torch implementation, for instance, behaves how I would expect:

import torch
import numpy as np 
import torchvision
x_np = np.ones((3, 64, 64)).astype(np.uint8)*255
x_torch = torch.tensor(x_np)
x_torch_jitter = torchvision.transforms.ColorJitter(0, 0, 0, 0.1)(x_torch)

#Jittered output is same as input image
(x_torch_jitter == 255).all() # is True

Environment

Running Ubuntu 20.04. cudatoolkit 11.6.2.

Installed cuCIM with pip.
python==3.8.13
cupy==11.2.0
cucim==22.10.00

(torch version for the above example == 1.13.0)

Other notes

• I came across this issue when using MONAI's RandCuCIM transform for data augmentation, as it calls cucim's color_jitter.
• I've tried a number of hue parameters -- anything non-zero seems to create the issue. Though the docstring seems to suggest anything between [-0.5, 0.5] is accepted, including negative values, negative values in fact return ValueError: If hue is a single number, it must be non negative. But I suppose that's a separate issue.

[benlansdell]

Looking at the CUDA kernel, here, it seems like the issue lies on line 208, here.

If s == 0 it sets the whole output to zero. But I think HSV is only black if v == 0.

Line 208 should be changed to:

    if (v == 0) {

??