Sigmoid as last layer

[laurentmih]

Hi, I just quickly wanted to say thanks: I used your structure to re-implement the model myself using resnet34 instead of VGG16, but turned to your code for inspiration or when I got stuck.

Anyways, I wanted to mention, you write here that you're not adding a Sigmoid at the end because your results converge to zero.
I ran into the same issue. The model learns within a single batch to only predict 0s. For me, the issue was that I had not added erosion to the matting stage. The result is that the "ground truth" for the model mostly contains 0s (>90%), causing it to predict 0s everywhere. I fixed it by also adding erosion, the same amount as dilation. This "balances" the labels to contain both 1s and 0s, punishing a model that only predicts 0s (the explanation here is a bit rubbish, happy to elaborate more if you want me to).

I saw here that you're only dilating, not eroding. Just wanted to suggest you check it out, maybe you're running into the same problem I had. Adding erosion might enable the usage of a sigmoid.

Thanks again for the inspiration!

[huochaitiantang]

Thanks for your mention! Although the eroded method was not mentioned in the paper "deep image matting", I also found that the unbalanced ground truth label generally exists. I conducted a new experience and got a better performance which is more closely to paper.