The wrong segmentation result in ''High Quality Segmentation..."

[Ianresearch]

Hi, I have tried to realize your method(High Quality Segmentation for Ultra High-resolution Images), it works well. But do you have met the problem that the wrong mask segmentation happens when the RGB-image area is similar to the real label area color. For example , I want to segment the sky area, but the white wall also to be thinked as sky.

[tcShen]

Do you mean the neighborhood regions with similar colors? Or the real color is similar to its neighbor region's annotation label color? Can you show your image and its segmentation results here? Thanks.

[Ianresearch]

The image is something like this:
https://small.justeasy.cn/wx/201801/9efb7f2097162a3595ce651a391d35ea.png
https://small.justeasy.cn/wx/201801/a4eb4e980da27692d1783af7b033285a.png
The whole wall is easy be segmented as sky, even the wall is far away form the sky in picture. What is more, the white car also easy be thinked as the sky, and the car thinked as not the sky before put into the mask refine net.
https://pic1.zhimg.com/v2-0f20b5c71062b2b03d1a483f2f27bef4_1440w.jpg?source=172ae18b

[tcShen]

Can you show me the coarse and the refined mask of the 3 images above? Thanks.