Model does not perform well on real world portrait images/videos

[anilsathyan7]

I trained the SINet model on EG1800 and the baidu augmented datasets, using the default settings in the code. I tried both cross entropy and lovasz loss functions for training. Both of them acheived mIOU ~ 94.5; but they do not seem to perform well on real world portrait images(even the provided SINet.pth checkpoint). I tried testing the model using the demo video from portrait-net repo and 'Visualize_video.py' script. In both cases there were artefacts on the images in the background regions and sometimes they appear on foreground regions also.

Here are the results:-

Output Videos:
test_results.zip

[dpcq]

do you try pretrained model?
can you show your Visualize_video code,because i meet some question when i run this code
thx!