These sections include a breif description of the usage of the code for the semantic segmnentation implementation.
The implementation uses a pre-trained version of the VGG19 network, available
from the loadcaffe package. The path
to the pre-trained network is specified inside model.lua.
The network is trained using the Cityscapes data
set. All images with fine annotations were renamed to the format
00XXXX.png, and corresponding lable files were created using these
convenient scripts. Be sure to set the correct path
variable inside data.lua:
dataPath = '/......../cityscapesProcessed/'The script main.lua contains useful parameters for the network training,
stored in table opt. It
also declares a high-level training function train().
To speciy if gradients should propagate through the entire network or only the
deconvolutional part of the network, use the function setParametersNet() with arguments
'model' or 'heads'.
File test.lua contains some useful functions for testing a trained network.