bad predictions about image-segmentation-keras

[louxy126]

[louxy126]

It‘s an image from the test floder, I trained the net with
parser.add_argument("--epochs", type = int, default = 100 ) parser.add_argument("--batch_size", type = int, default = 4 ) parser.add_argument("--val_batch_size", type = int, default = 4 ) parser.add_argument("--load_weights", type = str , default = "") parser.add_argument("--input_height", type=int , default = 224 ) parser.add_argument("--input_width", type=int , default = 224 )
and at the end of predict.py,i resize picture to 640,320
seg_img = cv2.resize(seg_img , (640 , 320 ))

[louxy126]

prediction with the same image use vgg_segnet
final outcome of train:
512/512 [==============================] - 37s 72ms/step - loss: 0.0897 - acc: 0.9318 - val_loss: 1.0100 - val_acc: 0.8151

[louxy126]

prediction with the same image use fcn32
final outcome of train:
512/512 [==============================] - 176s 345ms/step - loss: 1.4754 - acc: 0.4158 - val_loss: 1.6122 - val_acc: 0.3327