Why do I have grids in my segmentation mask?
I'm trying to train a model for semantic segmentation, by when I take a look a the resulting mask, I see some grid lines on it. What is the problem? How can I solve it? Here is what it looks like.
I'm trying to train a model for semantic segmentation, by when I take a look a the resulting mask, I see some grid lines on it. What is the problem? How can I solve it? Here is what it looks like.
Yea it's because of the rate of dilated convolution. It's high. There is a network called HDC (Hybrid Dilated Convolution), which helps you to cover all these gridding effects and fix them.
Here is the paper https://arxiv.org/pdf/1702.08502.pdf
Did you use dilated convolution? If yes, then the reason is the high value for dilation. If you remember how it works, you will realize that a high dilation rate will cause some gridding effects. Try to decrease the rate of it.