search

anilsathyan7 / Portrait-Segmentation

Real-time portrait segmentation for mobile devices
MIT License
638 stars 133 forks source link

about Boundary loss & Consistency Constraint Loss #8

Open huangxiaosa opened 4 years ago

huangxiaosa commented 4 years ago

Thanks for your work! I wonder if the bodundary loss & consistency constraint loss are included in the code? if not, could you please share more details to help me reproduce that?

anilsathyan7 commented 4 years ago

It's included in the portrait-net model (not in the uploaded version of code) and you can refer them through the following links:-

  1. https://github.com/dong-x16/PortraitNet
  2. https://www.sciencedirect.com/science/article/abs/pii/S0097849319300305

These additional techniques may not make much of a difference during inference (not easily noticed) at higher resolution on mobile phone, since we apply some post-processing anyway. In my experiments for portrait images, i did not find any real accuracy improvement on deployment(visually) after training on a sufficiently big dataset; except that the video became more stabilized due to temporal consistency.

I guess the only way to find out is to try different versions and compare them visually on application and experimentally with some metrics ....