search

xuebinqin / DIS

This is the repo for our new project Highly Accurate Dichotomous Image Segmentation
Apache License 2.0
2.22k stars 258 forks source link

the model starts to remove the object also when fine tuning on the isnet.pth weights #125

Open abhicake2048 opened 4 months ago

abhicake2048 commented 4 months ago

while training the model above isnet.pth weights i am getting skewed results as the number of epochs increase. the model starts to remove the foreground object from the images as the number of epochs increase. the training script is the original. i have uploaded the image in which the first image is the original image, the second image is the expected output and the last image is the output. as you can see that the trained model has removed the foreground object also. why is this happening? result6

Mlegbder commented 4 months ago

I'm also getting white or highlights keyed or transparent!

abhicake2048 commented 4 months ago

I'm also getting white or highlights keyed or transparent!

is it isnet_general_use.pth or isnet.pth??