What is the correct way to deal with the mean when fine-tuning a model?

[Mario Klingemann]

I am currently following the fine-tune tutorial to train the bvlc_reference_caffenet for different categories. These are not necessarily just normal photos but can contain also graphic elements which are higher in their average saturation or very biased towards certain colors, In general the approach already works quite nicely, but I am wondering if I can get even better results by tweaking certain parameters. One thing to change might be the mean_file which is currently using the default ilsvrc12/imagenet_mean. Would make any sense to replace that with a mean over my training set? And if yes, how do I calculate such a binaryproto file?

Mario Klingemann