This is a Fully Convolutional Networks(FCN), which is based on ResNet50, for Semantic Segmentation. I have fine tuned this net with VOC2012 train set, horizontal flip for data augmentation.
I tried to change the filter size in fc6 from 1*1 to 7*7 with padding 3. But there was no improvement. I guess the reason is that I reduced the number of filters in fc6 and fc7. I will not fix it, but you can still reference this code for your experiment.
fc6 layer with size 1*1: Google Drive
fc6 layer with size 7*7: Google Drive