Hi everyone

I am working on a project where I want to compare the performance of CNNs on RGB images and the converted greyscale images. In PyTorch there is a handy transform `torchvision.transforms.Grayscale(num_output_channels=1)`

to convert an RGB image into its greyscale version.

In the source code I found that they use the luma transform to do so: L = R * 299/1000 + G * 587/1000 + B * 114/1000. Essentially, this converts an RGB image into the Y plane of a YCbCr image (if I am not mistaken).

My question: do the mean and the standard deviation of a dataset of converted greyscale images change compared to the same dataset with RGB images? The mean should be the same right, since grey just means that R=G=B. What about the standard deviation?

Any help is very much appreciated!

All the best

snowe