How to calculate the IMG_MEAN?

[RaphaelDuan]

Dear @DrSleep : I have meet some problem when I use this project to train my dataset. It seems everything is work and I got a mIoU=0.945 on pascal 2012 ?! I use inference.py to test a image then I got a black image. Even I use one image to train, then I inference.py the image, the output image is a black one. Where is the problem? BTW, I regarded it maybe related to the IMG_MEAN at the head of train.py and inference.py. I did not find some code for calculating the IMG_MEAN in this project, if I should calculate it by myself what is the argument --is-training work for?