Fine tuning fully convolutional network(FCN)

[Bato]

Hi all, 

I'm trying to fine tuned FCN-8s to my own dataset. One thing that confused me a lot is in what way should we create the lmdb file. I understand that when fine tuning a CNN, we first need to create a text file with lines like "/path/to/image.png label". 

My questions are:

1. For FCN, since the label is pixel-wise, should we change the text file into: "/path/to/image.png /path/to/label_array"? This way, all the images and labels are embedded into one lmdb file. Is that correct? Or should we create two lmdb files for images and labels respectively? If that's the case, how can we change the "train.prototxt" to load two different lmdb files?

2. In what format should we create the label file? I heard some of the people use .png as well, but by doing this we need to somehow encoded the color of the label image. Can we just use an multiple dimension array as the label file? (Such as a .npy file or simply .cvs file to store the label matrix. )

Any suggestion would be welcome. Thank you!