serge-weihao / tmf-matting Goto Github PK

I am trying to use your pretrained model for image matting but my results are nowhere near as good as what you show in your publication. Would you mind taking a look and maybe provide some feedback on what I am doing wrong? It would be very much appreciated.

I am using your TMF_comp1k model checkpoint and running it either with our without TTA. This is the command with:

python demo/matting_demo.py \
    configs/mattors/gradloss/tmflaploss020tta8.py \
    ./TMF_Matting_comp1k.pth \
    ./bird.png \
    ./trimap.png \
    ./matte_tta.png

You can see my input images below as well as two images where I have used the alpha matte to substitute the background for a black, blue or white solid color. The first result is with TTA and the second is without.

The two results are pretty similar but neither gives me a very tight crop. What am I doing wrong?

Thank you in advance.

Would you mind sharing your trained checkpoints somewhere where I don't have to install a client to download it? Google Drive is pretty easy. Or just a web server or something.