AceCoooool / DSS-pytorch

:star: PyTorch implement of Deeply Supervised Salient Object Detection with Short Connection
MIT License
173 stars 53 forks source link

boundary noise #6

Open GuideWsp opened 6 years ago

GuideWsp commented 6 years ago

I train the DSS models, and test it. when i test it, i found that: the saliency object boundary have too much noise. do u have the same problem? maybe the reason is that: the code use mean to fuse, while the paper use a conv layer to fuse. i will try it later.

AceCoooool commented 5 years ago
  1. This is caused by “inference” (the z_fuse achieve here is a little different),the paper use z_fuse, z_2, z_3, z_4. You can see the output of demo (s-out6 and fusion without blur boundary)
  2. Current results is based on average fusion stragedy, I will train a learnable fusion soon(the paper using learnable way)