normalization using pre-trained vgg

Q: If I wanna extract features with VGG (pre-trained on the ImageNet) on my dataset, should I subtract the ImageNet mean, or should I calculate my dataset's mean and std firstly? Are there big differences between the ImageNet and other RGB datasets generally?

A:  Usually if your use case stays in the same data domain, the mean and std won’t be that different and you can try to use the ImageNet statistics.

I would recommend to use your own data statistics if you are dealing with another domain, e.g. medical images.

If you are in doubt that the ImageNet stats are a good fit for your data, I would re-calculate them on my dataset.

From: https://discuss.pytorch.org/t/about-normalization-using-pre-trained-vgg16-networks/23560