Car Window Refinement

[jamilkarabaev]

Do you think the model can learn this? I have 7k pieces.

Refining window segmentation of car(area of window that can be seen through the car)

I am showing image, then ground truth, then segmentation inputs.

cascade psp has to refine segmentation inputs to ground truth.

What do you think? possible, or too hard for model

First picture is logs while training. ious are zeroish... saw ur response on an issue saying its not good after 1400 its to one guy.