Training output perform much worse than pre-train segnet_basic_camvid model

[miketry0313]

Hi All,

I am fine tuning segnet_basic_camvid.caffemodel on Cityscape data by using segnet_basic_train.prototxt and segnet_basic_solver.prototxt. The output of loss and avg accuracy looks good (loss: 0.76, avg accuracy: 0.6). But when I using test_segmentation_camvid.py to display inference model result, it looks much worse than pre-train model. Can anyone give me some help? Thanks in advanced.

(Ground Truth)

( Pretrain Model Inference Result)

Below are the result of training output model before calculating bn statistics and after bn statistics.