Bad test

[Dawson-huang]

when I run:
python test_segmentation.py --model ENet/prototxts/enet_deploy_final.prototxt \ --weights ENet/enet_weights_zoo/cityscapes_weights.caffemodel \ --colours ENet/scripts/cityscapes19.png \ --input_image ENet/example_image/munich_000000_000019_leftImg8bit.png \ --out_dir ENet/example_image/
the result image is nice as your show!
But when I change the test image like this:

I get a bad result like this:

and others too:

So, like the topic that the model is a little bad to test, do you know the reason and guide me? Thanks.