foamliu/Deep-Image-Matting-PyTorch

Three changes compared to keras version v1

Closed this issue · 1 comments

rainsun1 commented

I noticed 3 insteresting changes in this version compared to v1:

  1. "compositional_loss" and "overall_loss" are removed, and only "alpha_prediction_loss" remains.
  2. "model refinement" step with dense layers or "torch.nn.linear" are removed, and only "encorder and decoder" layers remains.
    So, maybe this is a light version of v1.
  3. "generate_trimap(alpha)" function changes a lot, which can generate much large and usual unknown region size and is perhaps a key point for the improvement of the results.
foamliu commented

Thanks for the good summary.